2322 lines
119 KiB
Plaintext
2322 lines
119 KiB
Plaintext
|
|
|
|
===== 第 1 页 =====
|
|
O R A N G E B O O K
|
|
Hermes Agent
|
|
The Complete Guide
|
|
The first AI Agent that ships with reins built in
|
|
A practical guide to the Nous Research open-source
|
|
framework
|
|
The Agent That Grows With You
|
|
Keywords: Self-improving Agent · Cross-session Memory · Skill System · MCP · Multi-platform
|
|
For: Developers and AI enthusiasts who want to build a personal AI Agent
|
|
Version: v260408
|
|
HuaShu
|
|
WeChat: 花叔 · Bilibili: AI进化论-花生
|
|
Based on Hermes Agent v0.7.0. AI tools evolve rapidly — some content may change with future versions. Refer to official docs for the latest.
|
|
|
|
|
|
===== 第 2 页 =====
|
|
Contents
|
|
TABLE OF CONTENTS
|
|
Part 1: Concepts
|
|
§01
|
|
Not Another Agent: From Harness to Hermes
|
|
§02
|
|
Hermes Agent at a Glance: 60 Seconds to Understand
|
|
Part 2: Core Mechanisms
|
|
§03
|
|
The Learning Loop: An Agent That Builds Its Own Reins
|
|
§04
|
|
Three-Layer Memory: From Goldfish to Old Friend
|
|
§05
|
|
The Skill System: Self-Evolving Capabilities
|
|
§06
|
|
40+ Tools and MCP: Connect Everything
|
|
Part 3: Hands-On Setup
|
|
§07
|
|
Installation and Configuration: Three Ways
|
|
§08
|
|
First Conversation: Let Hermes Get to Know You
|
|
§09
|
|
Multi-Platform Access: Find It Anywhere
|
|
§10
|
|
Custom Skills: Teach Hermes New Tricks
|
|
§11
|
|
MCP Integration: Connect Your Tool Stack
|
|
Part 4: Real-World Scenarios
|
|
§12
|
|
Personal Knowledge Assistant: The Power of Cross-Session Memory
|
|
§13
|
|
Dev Automation: From Code Review to Deployment
|
|
§14
|
|
Content Creation: From Research to Draft
|
|
§15
|
|
Multi-Agent Orchestration: Run Three Horses at Once
|
|
Part 5: Deep Thinking
|
|
|
|
|
|
===== 第 3 页 =====
|
|
§16
|
|
Hermes vs OpenClaw vs Claude Code: Not a Choice
|
|
§17
|
|
The Boundaries of Self-Improving Agents: How Far Can It Go
|
|
|
|
|
|
===== 第 4 页 =====
|
|
§01 Not Another Agent: From Harness to Hermes
|
|
Not Another Agent: From Harness to Hermes
|
|
The OpenClaw craze hasn't even died down, and here comes Hermes Agent. It's not "yet another Agent tool"
|
|
— it's the first time the concept of Harness Engineering has been turned into a product.
|
|
Another one?
|
|
I get it. You're tired.
|
|
OpenClaw kicked off the lobster fever in late 2025. 26 million users. Every major tech company in China rushed to
|
|
build their own lobster-like product. Your social feeds were probably flooded with "I'm raising a lobster" posts for
|
|
weeks. And now, before the lobster craze has fully faded, something new pops up.
|
|
In February 2026, Nous Research released Hermes Agent. In under two months, GitHub stars shot past 27,000.
|
|
Your first reaction is probably: I still haven't figured out the lobster thing, and now this?
|
|
I spent a week tearing Hermes apart from top to bottom, and found that it's taking a completely different path
|
|
from OpenClaw. Hermes isn't another lobster. It's building something we've been talking about for a while but
|
|
nobody had actually shipped as a product.
|
|
What is Harness Engineering
|
|
If you've read the previous orange book Harness Engineering, feel free to skip this part. If not, here's the 30-second
|
|
version.
|
|
In early 2026, a consensus emerged in the AI coding world: the bottleneck isn't the model — it's the
|
|
environment. The LangChain team ran an experiment using the same model (GPT-5.2-Codex), only adjusting the
|
|
surrounding "harness" configuration. Scores jumped from 52.8% to 66.5%, rankings leaped from Top 30 to Top 5.
|
|
Not a single line of model code was changed.
|
|
Mitchell Hashimoto (creator of Terraform) was the first to name this: Harness Engineering. His approach was
|
|
straightforward — every time the AI made a mistake, he'd add a rule so it would never make the same mistake
|
|
again. The file was alive, always growing.
|
|
In that book, I broke the Harness down into five components. These five components are the key to understanding
|
|
Hermes.
|
|
The five-component mapping
|
|
Harness Engineering is a methodology — it tells you "what kind of harness you should build for your AI." But
|
|
methodologies have a problem: execution is entirely manual. You have to write CLAUDE.md yourself, configure
|
|
hooks yourself, build a memory system yourself, design workflows yourself.
|
|
|
|
|
|
===== 第 5 页 =====
|
|
What Hermes did is: build all five components in.
|
|
Harness
|
|
Component
|
|
Manual Implementation
|
|
Hermes Built-in System
|
|
Instruction Layer
|
|
Hand-write CLAUDE.md /
|
|
AGENTS.md
|
|
Skill system (markdown skill files, auto-created + self-
|
|
improving)
|
|
Constraint Layer
|
|
Configure hooks / linter / CI
|
|
Tool permissions + sandbox + toolset enabled on demand
|
|
Feedback Layer
|
|
Manual review / evaluator Agent
|
|
Self-improving Learning Loop (auto-retrospective after each
|
|
task)
|
|
Memory Layer
|
|
Manually maintain knowledge
|
|
base
|
|
Three-layer memory (session/persistent/Skill) + Honcho user
|
|
modeling
|
|
Orchestration
|
|
Layer
|
|
Build your own multi-Agent
|
|
pipeline
|
|
Sub-Agent delegation + cron scheduling
|
|
Look at the left column versus the right. The left side is all manual — you'd need to be an experienced engineer to
|
|
set it up. The right side is out of the box, ready the moment you install it.
|
|
This is the fundamental difference between Hermes and OpenClaw. OpenClaw gives you a configuration-as-
|
|
behavior system — you write a SOUL.md, and it becomes what you want. Its memory system is capable (Daily Logs
|
|
+ MEMORY.md + semantic search) and its Skill ecosystem is massive, but Skills are primarily written and
|
|
maintained by hand. Hermes has all five dimensions built in, and they run automatically.
|
|
Connecting the dots: If you've used Claude Code's CLAUDE.md + hooks + memory, you've already been doing
|
|
Harness Engineering manually. What Hermes does is turn that manual workflow into an automated system. From
|
|
"you build the harness for the AI" to "the AI builds its own harness."
|
|
Why Nous Research built this
|
|
The team behind Hermes Agent isn't a big company — it's an open-source AI research lab.
|
|
Nous Research has been described as "the mysterious force in the open-source community." The key figure,
|
|
Teknium, co-founded the lab and leads post-training. Early on they relied on Redmond AI for compute, and the
|
|
team has always been small. Yet the Hermes model family they produced (from Hermes 3's 8B/70B/405B to
|
|
Hermes 4's 14B/70B/405B) reached frontier-level performance through post-training alone. No pre-training from
|
|
scratch, no massive compute budget required.
|
|
This philosophy carries over to Hermes Agent: using open-source tools + any LLM API, even individuals can
|
|
deploy AI Agents that rival commercial solutions. MIT license, fully open source.
|
|
Their core principles are pretty clear. User control comes first. The model is steerable — you can adjust its behavior
|
|
as needed, free from corporate content policies. They explicitly don't do censorship, stating the model is
|
|
|
|
|
|
===== 第 6 页 =====
|
|
"unencumbered by censorship, neutrally aligned." At the same time, they don't compromise on creativity, math,
|
|
coding, or reasoning performance.
|
|
These principles shaped Hermes Agent's design philosophy: it doesn't make decisions for you — you set the
|
|
rules, it learns the rules, then it gets better and better. Control stays with the user; the system handles the
|
|
complexity of execution.
|
|
Not a replacement, but a progression
|
|
There's a common misconception: Hermes is here to replace Claude Code or OpenClaw. It's not. These three tools
|
|
solve problems at different levels.
|
|
Claude Code does interactive coding. You sit at the terminal, going back and forth, collaborating in real time. It's
|
|
your pair-programming partner.
|
|
OpenClaw does configuration-as-behavior. You write a SOUL.md, and it becomes what you want. Configuration is
|
|
transparent, the ecosystem is mature, with 5,700+ community Skills on ClawHub.
|
|
Hermes does autonomous background work + self-improvement. You don't need to sit beside it. It runs on its
|
|
own, learns on its own, evolves on its own. Online 24/7, reachable anytime through Telegram or Discord.
|
|
An interesting point: all three tools use the agentskills.io standard, so Skills are interoperable. A Skill you wrote in
|
|
Claude Code works in Hermes too, and vice versa. They're not three parallel lines — more like three roles with
|
|
different jobs in the same ecosystem.
|
|
核心建议
|
|
If you're just writing code, Claude Code is plenty. If you want a 24/7 background Agent that watches over your tasks
|
|
and gets smarter on its own, that's when you should look at Hermes.
|
|
Ships with the harness built in
|
|
Back to the question we started with: why is this time different?
|
|
The lobster craze taught everyday users something: AI can be "something you raise," not just "something you open
|
|
and use." OpenClaw's SOUL.md, memory system, and personalization let people experience for the first time what
|
|
"my AI" really feels like.
|
|
But lobster owners also discovered a problem: you have to build the entire harness yourself. Writing SOUL.md,
|
|
manually tweaking skills, periodically cleaning up memory. The lobster gets better with use, but only if you're
|
|
willing to invest the time feeding it.
|
|
Hermes takes a different approach: the harness comes welded into the factory settings, and it grows on its own.
|
|
From your very first conversation, Hermes starts automatically writing to memory, extracting Skills, and optimizing
|
|
workflows. The longer you use it, the deeper its understanding of you, the higher the quality of its work. You're not
|
|
training it — it's training itself.
|
|
Hermes is the first Agent that ships with the harness built in. And the harness grows on its own.
|
|
|
|
|
|
===== 第 7 页 =====
|
|
§02 Hermes Agent at a Glance: 60 Seconds
|
|
Hermes Agent at a Glance: 60 Seconds
|
|
One flow chart, one set of numbers, one comparison table. After these three, you'll know what Hermes is all
|
|
about.
|
|
Architecture in one picture
|
|
Hermes Agent's architecture can be strung together in a single line:
|
|
Learning Loop
|
|
→
|
|
Three-Layer Memory
|
|
→
|
|
Skill System
|
|
→
|
|
40+ Tools
|
|
→
|
|
Multi-Platform Gateway
|
|
From left to right, each module's job in one sentence:
|
|
The Learning Loop is Hermes's heart. After completing each task, it automatically does a retrospective: what
|
|
should be remembered, what Skills should be extracted, do existing Skills need optimization. This loop runs
|
|
continuously — you never have to trigger it manually.
|
|
Three-Layer Memory is Hermes's brain. Session memory remembers "what just happened," persistent memory
|
|
remembers "who you are and what you like," and Skill memory remembers "how to do things." Each layer has its
|
|
own job, stored in SQLite + FTS5 indexes, retrieved on demand rather than loaded all at once.
|
|
The Skill System is Hermes's skill library. Each Skill is a standalone markdown file, stored in the
|
|
~/.hermes/skills/ directory. There are three sources: bundled with the repo, created by the Agent itself, or
|
|
installed from the community Hub. The key feature: Skills aren't static — they self-improve through use.
|
|
40+ Built-in Tools are Hermes's hands and feet. Five categories: execution (run code, manipulate files),
|
|
information (search, scrape), media (images, video), memory (read/write storage), and coordination (sub-Agent
|
|
delegation). Plus MCP integration, connecting to 6,000+ external applications.
|
|
Multi-Platform Gateway is Hermes's front door. Telegram, Discord, Slack, WhatsApp, Signal, CLI — 14 platforms
|
|
supported. You can message it on Telegram, it processes in the background on your VPS, and conversations stay
|
|
continuous across platforms.
|
|
Key numbers
|
|
Version v0.7.0, released April 3, 2026. A few numbers worth noting:
|
|
|
|
|
|
===== 第 8 页 =====
|
|
Metric
|
|
Data
|
|
GitHub stars
|
|
27,000+ (two months after release)
|
|
First month growth
|
|
6,000+ stars
|
|
Built-in tools
|
|
40+
|
|
Supported platforms
|
|
14
|
|
MCP integrations
|
|
6,000+ apps
|
|
Sub-Agent concurrency
|
|
Up to 3
|
|
Minimum deployment cost
|
|
$5/month VPS
|
|
Memory usage
|
|
<500MB (without local LLM)
|
|
License
|
|
MIT (fully open source)
|
|
27,000+ stars in two months — that's fast. Keep in mind OpenClaw only reached its current scale thanks to the
|
|
lobster craze's social virality. Hermes hit these numbers without any comparable viral effect, which tells you the
|
|
developer community genuinely feels it solves a real problem.
|
|
Key differences from OpenClaw
|
|
After the lobster craze, the question everyone's asking is: what's actually different between Hermes and
|
|
OpenClaw?
|
|
|
|
|
|
===== 第 9 页 =====
|
|
Dimension
|
|
Hermes Agent
|
|
OpenClaw
|
|
Core
|
|
philosophy
|
|
Self-improving Learning Loop
|
|
Configuration-as-behavior (SOUL.md)
|
|
Memory
|
|
Three-layer self-improving
|
|
(session/persistent/Skill)
|
|
Multi-layer memory (Daily Logs/MEMORY.md/semantic
|
|
search), primarily manually maintained
|
|
Skill
|
|
maintenance
|
|
Agent auto-creates + self-improves
|
|
Manually written and maintained
|
|
User modeling
|
|
Honcho dialectical modeling (12-
|
|
layer identity inference)
|
|
Based on SOUL.md configuration
|
|
Multi-platform
|
|
access
|
|
14-platform Gateway
|
|
50+ messaging platforms
|
|
(Telegram/Discord/WhatsApp/Slack/Signal, etc.)
|
|
Ecosystem
|
|
scale
|
|
40+ built-in tools + MCP 6,000+
|
|
ClawHub 5,700+ community Skills
|
|
Deployment
|
|
Self-hosted (from $5 VPS)
|
|
Official hosting / self-hosted
|
|
Skill interop
|
|
Both use agentskills.io standard
|
|
The two biggest gaps are learning ability and user modeling. OpenClaw's Skills are primarily written and tuned
|
|
manually — their evolution depends on the community and active user maintenance. The longer you use Hermes,
|
|
the more precise its Skills become, the deeper its memory, the smoother its execution.
|
|
But OpenClaw has something Hermes can't match: ecosystem maturity. 5,700+ community Skills on ClawHub,
|
|
with ready-made solutions for all kinds of scenarios. Hermes's community is still in its early stages. The network
|
|
effects from 26 million users in the lobster craze gave OpenClaw a massive head start that technology alone can't
|
|
bridge.
|
|
One-line distinction: OpenClaw is a lobster you raise yourself. Hermes is a lobster that grows on its own. One
|
|
depends on your careful feeding; the other learns from its own experience.
|
|
$5 gets you up and running
|
|
Cost is something a lot of people care about. The answer might surprise you.
|
|
Hermes itself is free, MIT open source. You only pay for LLM API calls. Deployment cost depends on what setup you
|
|
choose:
|
|
Cheapest option: any $5/month VPS (Hetzner CX22 runs about $4/month; DigitalOcean, Vultr work too), Ubuntu
|
|
22.04. Without running a local LLM, memory usage stays under 500MB. Pair it with OpenRouter using Claude Haiku
|
|
or DeepSeek, and API costs stay low too.
|
|
|
|
|
|
===== 第 10 页 =====
|
|
Even cheaper: Serverless. Use Daytona or Modal as your backend — the environment hibernates when idle, wakes
|
|
up automatically when a message comes in. Between-session costs are practically zero.
|
|
Privacy-focused: Run Ollama on your VPS with a local 8B or 70B open-source model. API costs are zero, but you'll
|
|
need a beefier VPS (16GB+ RAM recommended).
|
|
No matter which option, a $5 VPS + Telegram Bot gives you a personal AI Agent running 24/7. That cost-
|
|
performance ratio beats subscription-based commercial Agent solutions by a mile.
|
|
核心建议
|
|
For comparison: Claude Code Pro subscription is $20/month, Max subscription is $200/month. With Hermes's $5
|
|
VPS + API cost setup, monthly expenses stay under $10 for most use cases. Of course, they're positioned differently
|
|
so a direct price comparison isn't entirely fair — but the barrier to entry is undeniably much lower.
|
|
Who is this for
|
|
One last question: is this book for you?
|
|
If any of the following applies, keep reading:
|
|
You've used Claude Code or OpenClaw and want an Agent that can run background tasks autonomously. Not the
|
|
kind where you sit and watch — the kind that keeps working while you sleep.
|
|
You know about Harness Engineering and you're curious to see what it looks like when the methodology
|
|
becomes a product.
|
|
You want to deploy a private AI Agent on your own VPS, where your data never leaves your server.
|
|
You're simply curious about what an open-source project that hit 27,000+ stars in two months actually got right.
|
|
In the following chapters, we'll start with the Learning Loop and peel back Hermes's core mechanisms layer by
|
|
layer.
|
|
|
|
|
|
===== 第 11 页 =====
|
|
§03 The Learning Loop: Self-Harnessing Agent
|
|
The Learning Loop: Self-Harnessing Agent
|
|
The most surprising thing about Hermes Agent isn't what it can do — it's that it changes. The more you use
|
|
it, the better it gets. This isn't marketing speak. It's an observable, verifiable closed-loop mechanism.
|
|
Starting with a real scenario
|
|
Say it's your first time asking Hermes to write a Python script. You tell it: write me a scraper that grabs titles and
|
|
summaries from a certain website.
|
|
It'll produce a working script. But the style might not be what you prefer, variable naming might not match your
|
|
conventions, error handling might not be the way you'd do it. Perfectly normal — it doesn't know you yet.
|
|
By the tenth time, everything's different. It knows you prefer httpx over requests. It knows you like writing error
|
|
logs to a file instead of printing to the terminal. It knows your project structure is typically organized by module
|
|
under a src/ directory. It knows you hate long functions.
|
|
Nobody taught it any of this. It figured it out on its own.
|
|
That's what the Learning Loop does.
|
|
Five steps, one closed loop
|
|
Hermes's Learning Loop has five steps. Individually, none of them are complicated, but strung together they form a
|
|
continuous improvement flywheel.
|
|
Curate Memory
|
|
→
|
|
Create Skill
|
|
→
|
|
Skill Self-Improvement
|
|
→
|
|
FTS5 Recall
|
|
→
|
|
User Modeling
|
|
Looks like five independent features? They actually have causal relationships. Memory provides the raw material
|
|
for Skill creation. Skills accumulate feedback during use, triggering self-improvement. FTS5 enables precise recall
|
|
of historical experience. User modeling assembles these fragments into a complete picture.
|
|
Let's break them down one by one.
|
|
Step one: Memory curation
|
|
After each conversation, Hermes actively decides what's worth remembering. Actively decides — not passively
|
|
stores.
|
|
|
|
|
|
===== 第 12 页 =====
|
|
Traditional conversation memory is brute-force: shove the entire chat history into the context window. The more
|
|
you chat, the longer the context, until it overflows. It's like a person trying to hold every experience they've ever
|
|
had in their mind at once. Normal people can't do that.
|
|
Hermes works more like a person writing a diary. After each conversation, it looks back: what was this about? Any
|
|
new discoveries? Did the user express any preferences? Then it writes what's worth keeping into a SQLite database,
|
|
with FTS5 full-text indexing.
|
|
The system also has a periodic nudge mechanism that reminds the Agent to review recent interactions. Kind of like
|
|
a journaling app sending you a notification: anything worth recording today?
|
|
Step two: Autonomous Skill creation
|
|
When Hermes finishes a reasonably complex task, it asks itself a question: will this solution be useful again in the
|
|
future?
|
|
If the answer is yes, it distills the solution into a Skill file and saves it to ~/.hermes/skills/ . This Skill is a
|
|
markdown file containing the task description, execution steps, and things to watch out for.
|
|
Here's an example: you ask Hermes to clean a CSV file and import it into a database. After it's done, it might create
|
|
a Skill called csv-to-database.md , recording common data-cleaning steps, your preferred database connection
|
|
method, and the field validation rules you typically need.
|
|
Next time you say "import this CSV for me," it doesn't start from scratch. It loads that Skill and follows the
|
|
approach you've already validated.
|
|
Step three: Skill self-improvement
|
|
Creating a Skill isn't the end. Every time it's used and you provide feedback, Hermes takes that feedback and
|
|
modifies the Skill itself.
|
|
Say you tell it "this import script should check if the table exists first." Hermes doesn't just add the check this one
|
|
time — it goes back and edits the Skill file, writing that rule in. Next time the Skill is used, the check is included by
|
|
default.
|
|
This process is a lot like continuous improvement in software development. Every time a production issue
|
|
happens, you don't just patch it — you update the documentation and standards too, preventing the same class of
|
|
problem from recurring.
|
|
Key distinction: Traditional AI tools have memory that's an accumulation of conversation logs. Hermes's memory
|
|
is a distillation of experience. One is a video tape, the other is a notebook. Video tapes keep getting longer until
|
|
they overflow; a notebook you can use indefinitely.
|
|
Step four: FTS5 cross-session recall
|
|
Remembering all this stuff is one thing. The real trick is finding the right piece at the right time.
|
|
|
|
|
|
===== 第 13 页 =====
|
|
Hermes uses SQLite's FTS5 extension for full-text indexing. Before each new conversation, it searches historical
|
|
memory based on the current topic and loads only the relevant parts into context. Not all history — just what's
|
|
needed.
|
|
This matters more than you'd think. Most AI tools either don't remember what you said last time or dump
|
|
everything in and slow to a crawl. Hermes's approach: ask a database question, it searches database-related
|
|
memories; ask a frontend question, it searches frontend memories. Like a well-organized note system with a table
|
|
of contents and index — you look up what you need.
|
|
FTS5 has another benefit: it's purely local. Your memory data doesn't need to be uploaded to any server — it's all
|
|
in a local SQLite file. When you move machines, just copy the ~/.hermes/ directory.
|
|
Step five: User modeling
|
|
The final step is Honcho user modeling, an optional external integration. What it does goes beyond remembering
|
|
what you said: it infers what kind of person you are.
|
|
After each conversation, Honcho analyzes the exchange and derives your preferences, habits, and goals. These
|
|
derivations aren't just records of what you said — they're deeper patterns generalized from your behavior.
|
|
For example, you never explicitly said "I prefer concise code style," but Honcho inferred it by analyzing the patterns
|
|
in how you modify code across multiple sessions. Next time it generates code, it defaults to the concise approach.
|
|
We'll dig deeper into Honcho's 12-layer identity modeling in the next chapter.
|
|
Mitchell Hashimoto's philosophy, automated
|
|
If you read the Harness Engineering orange book, you might remember Mitchell Hashimoto's approach. When
|
|
using Claude Code, he had a habit: every time the Agent made a mistake, he'd add a rule to CLAUDE.md.
|
|
"Don't use the any type in this project."
|
|
"Put test files in the __tests__ directory, not in src."
|
|
"Write commit messages in English, starting with a verb."
|
|
One rule at a time. After a few weeks, CLAUDE.md became an incredibly detailed project spec. The Agent went
|
|
from a clueless newcomer to a veteran who knew every unwritten rule of the project. Mitchell said it felt like
|
|
training a new team member.
|
|
What Hermes does is essentially the same thing, but automated.
|
|
You don't need to manually write CLAUDE.md. You don't need to summarize rules yourself after every mistake.
|
|
Hermes observes on its own, summarizes on its own, writes Skills on its own, and applies those rules on its own
|
|
the next time around.
|
|
|
|
|
|
===== 第 14 页 =====
|
|
Dimension
|
|
Mitchell's Way (Manual)
|
|
Hermes's Way (Automated)
|
|
Rule source
|
|
Human spots a problem, writes it down
|
|
Agent extracts from its own feedback
|
|
Storage location
|
|
CLAUDE.md (single file)
|
|
Multiple Skill files + memory database
|
|
Improvement trigger
|
|
Only when the human remembers to add a
|
|
rule
|
|
Automatic evaluation after every use
|
|
Cross-project
|
|
portability
|
|
Manually copy CLAUDE.md
|
|
Skills are global, shared across all projects
|
|
Improvement speed
|
|
Depends on how diligent the human is
|
|
Continuous and automatic — never gets
|
|
lazy
|
|
Of course, automated doesn't mean perfect. Mitchell's hand-written rules tend to be more precise, because
|
|
humans have a clearer understanding of their own needs. Hermes's auto-generated rules may need tweaking, may
|
|
misjudge. But here's the thing: it drops the barrier to zero. Not everyone has Mitchell's patience to maintain a
|
|
finely-tuned rule file. Hermes lets people who don't want to mess with configuration still enjoy the "gets better
|
|
with use" experience.
|
|
The flywheel acceleration effect
|
|
None of the five steps are individually novel. Memory, Skills, retrieval, user profiling — the AI field has seen all of
|
|
these before.
|
|
Hermes's innovation is wiring them into a closed loop. Memory feeds Skill creation. Skill usage generates new
|
|
memories. New memories trigger Skill improvement. Improved Skills produce better results. Better results make
|
|
user modeling more accurate. More accurate profiling makes the next round of memory curation more targeted.
|
|
This is a positive feedback loop. The more you use it, the stronger every step gets — and they get stronger
|
|
simultaneously. Like Amazon's flywheel: more users bring more data, more data brings better recommendations,
|
|
better recommendations bring more users.
|
|
The difference is that Hermes's flywheel spins for a single user. It doesn't need data from millions of users to
|
|
improve — just your own usage history. Use it for three to five days, and you'll notice a clear difference.
|
|
核心建议
|
|
The Learning Loop's effectiveness is directly tied to how often you use it. If you only use it once or twice a week,
|
|
improvement will be slow. But if you treat it as your daily work partner, using it every day, the flywheel spins
|
|
remarkably fast. That's why Hermes is particularly well-suited for deploying on a $5 VPS running 24/7 — let it keep
|
|
accumulating.
|
|
What this means
|
|
Back to the title: the Agent builds its own harness.
|
|
|
|
|
|
===== 第 15 页 =====
|
|
In Harness Engineering, the harness is human-made. You write CLAUDE.md, configure hooks, design feedback
|
|
mechanisms. All of that requires ongoing human investment.
|
|
Hermes's approach: let the Agent weave its own harness while it runs. When it veers off course, it self-corrects
|
|
and remembers the lesson. When it finds a good method, it distills it into a Skill for future reuse. When it
|
|
encounters a new user, it builds its own understanding model.
|
|
This doesn't mean humans are out of the picture. You can manually edit Skill files anytime, delete inappropriate
|
|
memories, adjust the user profile. But by default, the system is self-driven.
|
|
Next chapter, we look at the most critical infrastructure in this loop: the three-layer memory system. If the Learning
|
|
Loop is the engine, memory is the fuel.
|
|
|
|
|
|
===== 第 16 页 =====
|
|
§04 Three-Layer Memory: From Goldfish to Old Friend
|
|
Three-Layer Memory: From Goldfish to Old Friend
|
|
Most AI chat tools have the memory of a goldfish — whatever was said last round is forgotten by the next.
|
|
Hermes aims to be an old friend: one who remembers what you said, knows what kind of person you are,
|
|
and has learned how you like to get things done.
|
|
Why memory is the hardest problem
|
|
You might think AI memory is just saving chat logs. Store them, load them next time, done.
|
|
Not that simple. An active user chats thousands of words a day with their AI. That's tens of thousands per month.
|
|
Cram it all into the context window and either it won't fit, or the model gets sluggish from information overload.
|
|
And most of a chat log is noise and repetition — the actually valuable information might be just 10%.
|
|
Good memory isn't about storing more — it's about finding what matters.
|
|
Hermes solves this with a three-layer architecture, each layer handling a different type of memory.
|
|
Layer one: Session memory
|
|
Session memory answers the question: what happened?
|
|
Every conversation's content, tool calls, and return results are written to a SQLite database with FTS5 full-text
|
|
search indexing. This is episodic memory, analogous to the hippocampus in the human brain.
|
|
The key design decision is on-demand retrieval rather than loading everything. When a new conversation starts,
|
|
Hermes doesn't stuff all past conversation history into context. It searches relevant historical fragments using FTS5
|
|
based on the current topic, loading only what's needed.
|
|
The benefits of this approach:
|
|
Approach
|
|
Load Everything
|
|
On-Demand Retrieval (Hermes)
|
|
Context usage
|
|
Grows linearly with conversation volume
|
|
Essentially constant
|
|
Retrieval precision
|
|
Everything's there, but nothing's findable
|
|
Precise keyword matching
|
|
Long-term viability
|
|
Hits the wall after a few days
|
|
Works for months, even years
|
|
Response speed
|
|
Gets slower over time
|
|
Stays essentially the same
|
|
FTS5 is SQLite's full-text search extension — no extra database installation needed. All data lives in local SQLite
|
|
files, no network dependency, no privacy concerns. Your conversation memory never leaves your machine.
|
|
|
|
|
|
===== 第 17 页 =====
|
|
Layer two: Persistent memory
|
|
Persistent memory answers the question: who are you?
|
|
This layer doesn't store conversation content — it stores durable state distilled from conversations. Things like
|
|
your coding preferences, project structure habits, commonly used toolchain, work schedule patterns. These persist
|
|
across sessions and don't disappear when you start a new conversation.
|
|
Technically, persistent memory is also stored in SQLite, managed through the memory tool. The entire solution is
|
|
file-level: no external servers, no cloud sync, data lives in the ~/.hermes/ directory.
|
|
This means you can:
|
|
Back up ~/.hermes/ to a USB drive and continue on a different machine
|
|
When deploying with Docker, mount the /opt/data directory to the host to persist state
|
|
Sync across devices using cloud storage (real-time SQLite sync isn't recommended, but periodic copying works
|
|
fine)
|
|
Portability is an underrated feature. Many AI tools lock your memory in the cloud — switch tools and you start
|
|
from zero. Hermes's memory is your own files, portable however you want.
|
|
Layer three: Skill memory
|
|
Skill memory answers the question: how to do things?
|
|
The first two layers remember what happened and who you are. The third layer remembers methodologies and
|
|
operating procedures. Each Skill is a markdown file in ~/.hermes/skills/ , human-readable and editable.
|
|
These three layers correspond to three types of memory in cognitive science:
|
|
Episodic Memory
|
|
What happened
|
|
→
|
|
Semantic Memory
|
|
What the world is like
|
|
→
|
|
Procedural Memory
|
|
How to do things
|
|
Learning to ride a bicycle is all three layers working together: you remember falling last time (episodic), you know
|
|
to keep your center of gravity low (semantic), and your body automatically balances (procedural). Hermes handles
|
|
tasks the same way: it remembers how you edited code last time, knows your preferences, and has a proven
|
|
execution plan at hand.
|
|
Three layers in action: You say "help me deploy this project." Hermes first searches session memory with FTS5,
|
|
finding the port conflict you hit during your last deployment (episodic). Then it checks persistent memory,
|
|
knowing you use Alibaba Cloud ECS with Nginx reverse proxy (semantic). Finally it loads the deployment-checklist
|
|
Skill and follows the steps you've already validated (procedural). Each layer doing its job.
|
|
|
|
|
|
===== 第 18 页 =====
|
|
Honcho: Knows you better than you know yourself
|
|
Beyond the three local memory layers, Hermes has an optional add-on: the Honcho user modeling system,
|
|
developed by Plastic Labs.
|
|
What Honcho does goes a step deeper than remembering what you said. It infers your characteristics as a person.
|
|
The official term is dialectical user modeling, covering 12 identity layers.
|
|
What do these 12 layers actually mean? Let's use a scenario to illustrate.
|
|
Say you've been asking Hermes to write Python scripts every day for three straight weeks. During this time,
|
|
Honcho might infer:
|
|
Technical level: You're not a complete beginner, but not an expert either. You can read code but struggle to
|
|
write it from scratch
|
|
Work rhythm: You're usually active from 9-11 PM, likely a personal project after work
|
|
Communication style: You prefer seeing results first and asking about principles later; you don't like long
|
|
explanations
|
|
Goal inference: You're probably working on a data analysis project, since recent tasks all revolve around data
|
|
processing
|
|
Emotional patterns: You get a bit frustrated when code throws errors; concise, direct answers work better in
|
|
those moments
|
|
Preference contradictions: You say you want comprehensive comments, but when you actually review code
|
|
you never read them
|
|
Notice that last one. Honcho catches inconsistencies between what you say and what you do. Your stated
|
|
preferences and your revealed preferences may differ — dialectical modeling pays attention to both.
|
|
These inferences get injected into subsequent conversation prompts as invisible context. You don't see this
|
|
information, but you feel Hermes becoming more attuned to you.
|
|
Compared to Claude Code's auto-memory
|
|
Claude Code also has a memory system: CLAUDE.md files and auto-memory. How do they compare with Hermes's
|
|
memory?
|
|
|
|
|
|
===== 第 19 页 =====
|
|
Dimension
|
|
Claude Code
|
|
Hermes Agent
|
|
Memory format
|
|
CLAUDE.md + auto-memory text files
|
|
SQLite database + FTS5 index + Skill
|
|
files
|
|
Write method
|
|
CLAUDE.md manually written, auto-memory semi-
|
|
automatic
|
|
Fully automatic, human can override
|
|
anytime
|
|
Retrieval method
|
|
CLAUDE.md loaded in full at startup
|
|
On-demand FTS5 full-text search
|
|
Memory
|
|
granularity
|
|
Project-level (one CLAUDE.md per project)
|
|
Both global-level and project-level
|
|
User modeling
|
|
None (user writes their own preferences)
|
|
Honcho auto-infers user profile
|
|
Procedural
|
|
memory
|
|
Instructions in CLAUDE.md
|
|
Standalone Skill files, self-improving
|
|
Cross-project
|
|
sharing
|
|
~/.claude/CLAUDE.md (global instruction file)
|
|
All memory is inherently global
|
|
Storage limit
|
|
CLAUDE.md recommended at a few KB
|
|
SQLite's theoretical limit is very high
|
|
The two have different design philosophies. Claude Code's CLAUDE.md follows a human-writes, AI-executes model
|
|
— the upside is full human control, the downside is it requires ongoing maintenance. Hermes follows an AI-writes,
|
|
human-reviews model — the upside is low barrier and high automation, the downside is auto-generated content
|
|
may not always be accurate.
|
|
Which is better depends on the use case. If you're a heavy Claude Code user who's spent weeks carefully crafting
|
|
your CLAUDE.md, your hand-woven harness may be more precise than what Hermes auto-generates. But if you
|
|
don't want to spend time maintaining config files, Hermes's fully automated approach is genuinely easier.
|
|
Memory isn't a silver bullet
|
|
After all the good stuff, let's also talk about limitations.
|
|
What should be remembered:
|
|
User preferences and habits (code style, tool choices, communication style)
|
|
Project context (architecture decisions, tech stack, file structure)
|
|
Validated solutions (Skills)
|
|
Recurring patterns (like how to handle certain classes of errors)
|
|
What shouldn't be remembered:
|
|
One-off task details (write me a birthday greeting — no need to remember that)
|
|
Outdated information (an API version number from three months ago that's probably changed)
|
|
Wrong inferences (Hermes may misjudge your preferences — these should be cleaned up)
|
|
|
|
|
|
===== 第 20 页 =====
|
|
Sensitive information (passwords, keys, personal identity info shouldn't enter the memory store)
|
|
注意
|
|
Hermes's memory system currently has no automatic expiration mechanism. With long-term use, the memory
|
|
database will keep growing. It's a good idea to periodically check the size of the ~/.hermes/ directory and clean
|
|
up outdated Skill files. This is a known area for improvement.
|
|
There's also the problem of memory pollution. If Hermes remembers incorrect information from early
|
|
conversations, that error can persist and affect later behavior. For instance, if it mistakenly remembers you prefer
|
|
Python 2, subsequent generated code might carry Python 2 syntax.
|
|
So periodic memory audits are worth doing. Check which Skills are in ~/.hermes/skills/ , delete ones that
|
|
don't fit. Review persistent memory, correct wrong inferences. Like tidying up a notebook — flipping through it
|
|
occasionally, you'll find plenty that needs updating.
|
|
From storage to understanding
|
|
Back to the title: from goldfish to old friend.
|
|
A goldfish's problem isn't that it has no eyes — it's that it has no memory. Every time it sees you is like the first
|
|
time. Most AI tools are like this: open a new conversation, and everything starts from scratch.
|
|
An old friend is different. An old friend knows your temper, your habits, knows that when you say you don't care,
|
|
you actually do. An old friend doesn't need you to explain the backstory every time, because the backstory is
|
|
already there.
|
|
Hermes's three-layer memory plus Honcho user modeling is walking this path. Session memory provides the raw
|
|
material, persistent memory provides the profile, Skill memory provides the methodology, and Honcho assembles
|
|
these fragments into a complete understanding of you.
|
|
The longer you use it, the deeper the understanding. This isn't a slogan — it's a mathematical inevitability of
|
|
three-layer memory plus a closed-loop learning system.
|
|
Next chapter, we look at another core mechanism in Hermes: the Skill system. If memory is "knowing," Skills are
|
|
"being able to do."
|
|
|
|
|
|
===== 第 21 页 =====
|
|
§05 The Skill System: Capabilities That Evolve on Their
|
|
Own
|
|
Skills: Self-Evolving Capabilities
|
|
OpenClaw's Skills require you to write and maintain them by hand. Hermes' Skills grow on their own and get
|
|
better over time. This difference defines two entirely different user experiences.
|
|
What Exactly Is a Skill
|
|
In Hermes, each Skill is a standalone markdown file stored in the ~/.hermes/skills/ directory. It captures the
|
|
agent's procedural memory for how to do something.
|
|
Think of it this way: you're training a new colleague to write weekly reports. The first time, you walk them through
|
|
every step. The second time, they still ask a few questions. By the third time, they've got it. A Skill is that "after the
|
|
third time" state — the agent solidifies the method into a reusable document.
|
|
Skills come from three sources:
|
|
Source
|
|
Description
|
|
Scale
|
|
Bundled
|
|
Skills
|
|
Pre-built capabilities that ship with the install, covering common scenarios like
|
|
MLOps, GitHub workflows, and research
|
|
40+
|
|
Agent-
|
|
Created
|
|
After completing complex tasks, the agent automatically distills solutions into
|
|
Skills
|
|
Grows with usage
|
|
Skills Hub
|
|
Community-contributed skill packs, installable with one click
|
|
Continuously
|
|
growing
|
|
These three sources aren't equal. Bundled Skills are the starting point, Skills Hub is the accelerator, but agent-
|
|
created Skills are Hermes' real killer feature.
|
|
agentskills.io: A Universal Language for Skills
|
|
Hermes' Skills aren't a walled garden. They follow the agentskills.io standard, which is already supported by 30+
|
|
tools including Claude Code, Cursor, Copilot, Codex CLI, and Gemini CLI.
|
|
Skills you wrote for Claude Code work directly in Hermes. And vice versa.
|
|
This is different from the App Store model. App Stores mean one ecosystem per platform, forcing developers to
|
|
build for each one. agentskills.io is more like a USB port — one Skill plugs in anywhere and just works.
|
|
|
|
|
|
===== 第 22 页 =====
|
|
For people already using Claude Code, your accumulated Skill assets aren't locked into any single tool. Switch to
|
|
Hermes, or use both simultaneously — Skills migrate seamlessly.
|
|
|
|
|
|
===== 第 23 页 =====
|
|
Self-Improvement: Skills Get Better With Use
|
|
This is the biggest difference between Hermes and every other agent Skill system out there.
|
|
Traditional Skills require manual maintenance. You write a code review Skill, and it follows your steps exactly.
|
|
Discover that a certain step doesn't work well in practice? You have to go in and fix it manually. That's primarily
|
|
how OpenClaw's 5,700+ community Skills operate.
|
|
Hermes' Skills are alive. They run inside a learning loop, automatically optimizing based on real feedback.
|
|
Here's the mechanism:
|
|
1
|
|
Execute the Skill
|
|
The agent follows the steps recorded in the Skill to complete the task
|
|
2
|
|
Collect Feedback
|
|
The user's reactions (satisfied / unsatisfied / corrections) get logged into session memory
|
|
3
|
|
Update the Skill
|
|
The agent analyzes feedback and automatically modifies the relevant steps in the Skill file
|
|
4
|
|
Next Execution Uses the New Version
|
|
The improved Skill takes effect automatically in subsequent tasks
|
|
Sounds idealistic? It is, to some extent — results depend on the LLM's capabilities and feedback quality. But the
|
|
direction is right: let the agent learn from experience instead of waiting for someone to maintain it.
|
|
Comparison with Mitchell Hashimoto: Mitchell said that when using Claude Code, he "adds a rule to CLAUDE.md
|
|
every time it makes a mistake." Hermes automates that process. You don't have to add rules yourself — the agent
|
|
observes, summarizes, and writes them into Skills. The trade-off is you give up some control over the rules.
|
|
Key Differences from OpenClaw Skills
|
|
OpenClaw's ClawHub has 5,700+ community Skills, far outnumbering Hermes. But the design philosophies are
|
|
completely different:
|
|
|
|
|
|
===== 第 24 页 =====
|
|
Dimension
|
|
OpenClaw Skills
|
|
Hermes Skills
|
|
Creation
|
|
Manually written SOUL.md
|
|
Agent-created + manually written
|
|
Maintenance
|
|
Manual updates
|
|
Auto-evolution + manual intervention
|
|
Personalization
|
|
Generic templates, fork and customize
|
|
Grows organically from your usage habits
|
|
Interoperability
|
|
agentskills.io standard
|
|
agentskills.io standard (interoperable)
|
|
Ecosystem Size
|
|
5,700+ (large)
|
|
40+ bundled + community (growing)
|
|
OpenClaw's strength is scale and transparency. One look at the SOUL.md tells you exactly what a Skill does —
|
|
every step is something you wrote, so you know what's going on.
|
|
Hermes' strength is adaptability. The same "write code" Skill, after three weeks of use by a Python developer and
|
|
a Rust developer, will have evolved into two completely different versions. Not a generic template — a custom fit.
|
|
The two aren't mutually exclusive. The agentskills.io standard lets Skills move between them. You can absolutely
|
|
install a Skill from ClawHub into Hermes, then let Hermes continuously improve it through use.
|
|
|
|
|
|
===== 第 25 页 =====
|
|
In Practice: Letting Hermes Create a Skill on Its Own
|
|
Enough about mechanisms — let's look at a concrete example.
|
|
Say every morning you need Hermes to sort through yesterday's GitHub notifications, summarizing important PRs
|
|
and Issues. The first few times, you have to spell out the request each time:
|
|
核心建议
|
|
"Go through my GitHub notifications from yesterday, sort by importance, separate PRs and Issues, and ignore
|
|
automated bot notifications."
|
|
After the third or fourth time, Hermes does something in the background: it distills this recurring task pattern into a
|
|
Skill file. You'll find a new markdown file in ~/.hermes/skills/ , looking something like this:
|
|
# GitHub Daily Digest
|
|
## Trigger Conditions
|
|
User mentions "GitHub notifications", "daily summary", etc.
|
|
## Steps
|
|
1. Call GitHub MCP to fetch notifications from the past 24 hours
|
|
2. Filter out automated notifications from bot accounts
|
|
3. Group by type (PR / Issue / Discussion)
|
|
4. Sort by importance (mention > review request > other)
|
|
5. Present as a concise list
|
|
## User Preferences
|
|
- Only titles and status needed, no detailed content
|
|
- PRs and Issues displayed separately
|
|
From this point on, you just say "check GitHub" and Hermes knows what to do.
|
|
What's even more interesting is what happens next. One day you say "include Discussions this time too." Hermes
|
|
doesn't just add them for this one request — it updates the rules in the Skill file. Next time, it'll include
|
|
Discussions even without you asking.
|
|
That's what "self-evolution" actually means in practice. No mysterious AI breakthrough — just an automated loop
|
|
of "user corrects -> rules update -> next execution applies the change."
|
|
注意
|
|
Skill self-improvement requires your feedback to be clear enough. If you just feel "something's off" but don't say
|
|
what specifically, the agent can't improve accurately. Good feedback = good evolution direction.
|
|
|
|
|
|
===== 第 26 页 =====
|
|
§06 40+ Tools & MCP: Connect Everything
|
|
40+ Tools & MCP: Connect Everything
|
|
No matter how smart an agent is, it can't get real work done without tools. Hermes ships with 40+ built-in
|
|
tools covering everything from running code to sending messages. MCP extends its reach to 6,000+ external
|
|
applications.
|
|
Five Categories at a Glance
|
|
Hermes' tools fall into five categories. You don't need to memorize them all — just know these capabilities exist
|
|
and look them up when you need them.
|
|
Category
|
|
Core Tools
|
|
What They Do
|
|
Execution
|
|
terminal, code_execution,
|
|
file
|
|
Run commands, execute code (sandboxed), read/write files
|
|
Information
|
|
web, browser,
|
|
session_search
|
|
Web search, browser automation, search conversation history
|
|
Media
|
|
vision, image_gen, tts
|
|
Understand images, generate images, text-to-speech
|
|
Memory
|
|
memory, skills, todo, cronjob
|
|
Operate the memory layer, manage Skills, task planning, scheduled
|
|
jobs
|
|
Coordination
|
|
delegation, moa, clarify
|
|
Delegate to sub-agents, multi-model reasoning, ask user for
|
|
clarification
|
|
A few worth calling out individually:
|
|
session_search is a fairly unique Hermes capability. It uses FTS5 full-text indexing to search conversation history,
|
|
paired with LLM summarization, letting the agent quickly recall "that approach we discussed last week." Most
|
|
agents don't have this — every conversation starts from scratch.
|
|
moa (Multi-model Orchestrated Answering) lets Hermes call multiple LLMs simultaneously, synthesizing their
|
|
responses into a final answer. Useful for scenarios requiring high reliability, like fact-checking or technical
|
|
decisions.
|
|
cronjob defines scheduled tasks using natural language. Say "check my GitHub notifications every morning at
|
|
9am" and it creates a timed trigger. No cron expressions, no scheduler configuration needed.
|
|
|
|
|
|
===== 第 27 页 =====
|
|
Toolsets: Not Everything On, but On-Demand
|
|
Having all 40+ tools enabled at once doesn't make sense. An agent helping you write code doesn't need Home
|
|
Assistant permissions, and a calendar-managing agent has no use for code_execution.
|
|
Hermes solves this with the Toolset mechanism. Tools are grouped by function and enabled or disabled in
|
|
config.yaml as needed:
|
|
# config.yaml example
|
|
toolsets:
|
|
- web # web search
|
|
- terminal # terminal commands
|
|
- file # file operations
|
|
- skills # Skill management
|
|
- delegation # sub-agent delegation
|
|
# - homeassistant # comment out what you don't need
|
|
# - rl # reinforcement learning, most people won't need this
|
|
This isn't just a feature toggle. Fewer enabled tools means the agent stays more focused, responds faster, and
|
|
consumes fewer tokens. If all you need is a file-organizing assistant, enabling just the file and memory Toolsets is
|
|
enough.
|
|
Toolsets also serve as security boundaries. The constraint layer mentioned in S03 is implemented at the tool level
|
|
through Toolsets. You get precise control over what the agent can and can't touch.
|
|
MCP: A Unified Interface to 6,000+ Apps
|
|
The 40+ built-in tools cover general scenarios. But everyone's toolchain is different — you might use Jira for project
|
|
management, Notion for docs, Slack for communication. How do you get Hermes to work with these?
|
|
MCP (Model Context Protocol).
|
|
MCP is an open protocol that defines a communication standard between AI agents and external tools. Hermes
|
|
supports connecting to any MCP Server via stdio or HTTP. The MCP ecosystem currently covers 6,000+ applications
|
|
— GitHub, Slack, Jira, Google Drive, databases, you name it.
|
|
Integration is straightforward — add a config block to config.yaml:
|
|
# Connect to GitHub MCP Server
|
|
mcp_servers:
|
|
github:
|
|
command: npx
|
|
args: ["-y", "@modelcontextprotocol/server-github"]
|
|
env:
|
|
GITHUB_PERSONAL_ACCESS_TOKEN: ${GITHUB_TOKEN}
|
|
Once configured, Hermes can use GitHub's capabilities: create Issues, review PRs, check repo status. No code to
|
|
write, no custom tools to build — MCP Servers are plug-and-play capability extensions.
|
|
|
|
|
|
===== 第 28 页 =====
|
|
Hermes also supports per-server tool filtering. An MCP Server might expose 20 tools, but you might only want the
|
|
agent to use 3 of them. Tool filtering gives you precise control over the capabilities exposed to the agent.
|
|
|
|
|
|
===== 第 29 页 =====
|
|
Sub-Agent Delegation: Three Horses Running at Once
|
|
Delegation is Hermes' most powerful coordination tool. It can spawn sub-agent instances, distributing tasks for
|
|
parallel execution.
|
|
Independent context. Each sub-agent has its own conversation context, isolated from the others. The main agent
|
|
passes in necessary background when assigning tasks, and each sub-agent works in its own space.
|
|
Restricted toolsets. You can specify which tools each sub-agent can use. The research one only needs web and
|
|
browser; the coding one only needs terminal and file. This is both an efficiency optimization and a security
|
|
measure.
|
|
Max 3 concurrent. This limit is intentional. Three concurrent sub-agents already cover most parallel scenarios
|
|
(research, coding, and testing simultaneously), and more would be too hard to coordinate effectively.
|
|
Main Agent
|
|
Decompose tasks
|
|
→
|
|
Sub-Agent A
|
|
Research
|
|
→
|
|
Sub-Agent B
|
|
Code
|
|
→
|
|
Sub-Agent C
|
|
Test
|
|
If you've used Claude Code's multi-instance parallelism, this will feel familiar. The difference is that Claude Code's
|
|
multiple instances are manually opened by you, with no coordination between them. Hermes' sub-agent
|
|
delegation is the agent autonomously deciding when to distribute tasks and how to consolidate results.
|
|
Real-world feel: Sub-agents work best for "do several unrelated things and then combine the results." For
|
|
example, ask Hermes to write a technical blog post and it might dispatch one sub-agent to research the latest
|
|
materials, another to analyze competitor articles, and a third to organize code examples — then the main agent
|
|
integrates all three into a first draft.
|
|
Tool Permissions & Sandboxing: The Constraint Layer in Action
|
|
The previous chapters covered the learning loop, memory, and Skills — all mechanisms for making the agent more
|
|
capable. But the more capable an agent gets, the more constraints matter.
|
|
Hermes implements three layers of constraints at the tool level:
|
|
Toolset control. Only tools enabled in config.yaml can be called by the agent — the coarsest-grained switch.
|
|
code_execution sandbox. Code runs in an isolated environment, separate from your system. Even if the agent
|
|
executes problematic code, it won't affect your filesystem.
|
|
Sub-agent restricted toolsets. When delegating to a sub-agent, you can specify a subset of tools it can use. A sub-
|
|
agent responsible for searching doesn't need — and shouldn't have — terminal access.
|
|
If you've read the Harness Engineering orange book, you'll recognize that this is exactly how the constraint layer
|
|
(hooks/linter) is implemented in Hermes. Same principle: give the agent enough capability to complete tasks,
|
|
but don't hand over unnecessary permissions.
|
|
|
|
|
|
===== 第 30 页 =====
|
|
核心建议
|
|
For security-sensitive scenarios (like running on production servers), enable only the necessary Toolsets and use
|
|
MCP's per-server filtering to further restrict the tools exposed to the agent. Better to have the agent ask "I need XX
|
|
permission" than to leave everything open by default.
|
|
These three layers of constraints together form a pragmatic security model. It doesn't chase theoretically perfect
|
|
isolation — instead, it finds a balance between usability and security. The more trust you give the agent, the more
|
|
it can do. But even with default settings unchanged, Hermes won't perform dangerous operations without your
|
|
knowledge.
|
|
|
|
|
|
===== 第 31 页 =====
|
|
§07 Installation & Configuration: Three Approaches
|
|
Installation & Configuration
|
|
From zero to running in as little as 5 minutes. This section covers three installation methods — from local
|
|
development to a 24/7 server. Pick the one that fits.
|
|
Option 1: Local Install (5 Minutes to Get Started)
|
|
Local installation is the most straightforward, perfect for people who want to try it before committing to running it
|
|
long-term. The only prerequisite is having git on your machine.
|
|
1
|
|
Run the One-Line Installer
|
|
Open your terminal and paste this:
|
|
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
|
|
The installer automatically handles Python, Node.js, and all dependencies. Works on macOS, Linux, and WSL2.
|
|
2
|
|
Configure the LLM Backend
|
|
After installation, edit the config file:
|
|
# Config file location
|
|
~/.hermes/config.yaml
|
|
Enter your model API key (we'll cover how to choose a model shortly).
|
|
3
|
|
Launch Hermes
|
|
hermes
|
|
That's it — one word. If you see the welcome message, you're good to go.
|
|
核心建议
|
|
If you manage Python with uv, you can also clone the repo and install via uv pip install -e ".[all]" . Same
|
|
result — use whichever you prefer.
|
|
Option 2: Docker (Clean Isolation)
|
|
Don't want to install a bunch of dependencies on your machine? Docker is the cleanest choice.
|
|
|
|
|
|
===== 第 32 页 =====
|
|
docker pull nousresearch/hermes-agent:latest
|
|
docker run -v ~/.hermes:/opt/data nousresearch/hermes-agent:latest
|
|
Key flag: -v ~/.hermes:/opt/data maps the container's data volume to your host machine. All of Hermes' state
|
|
(memory, Skills, config) lives in /opt/data — one single directory. Delete and rebuild the container, your data
|
|
survives.
|
|
This is a nice design choice. Unlike some tools that scatter state across various paths, everything in Hermes lives
|
|
under ~/.hermes/ . When it's time to migrate, just pack up that directory.
|
|
Option 3: $5 VPS for 24/7 Uptime
|
|
If you want Hermes always online, independent of whether your computer is running, a $5/month VPS is all you
|
|
need.
|
|
Recommended setups:
|
|
VPS Provider
|
|
Monthly Cost
|
|
Notes
|
|
Hetzner CX22
|
|
~$4/mo
|
|
Best value, European nodes
|
|
DigitalOcean Droplet
|
|
$5/mo
|
|
Singapore/US West nodes
|
|
Vultr
|
|
$5/mo
|
|
Tokyo node, low latency
|
|
Pick Ubuntu 22.04 LTS, SSH in, run the install script — identical to local installation. If you're not running local
|
|
models, memory usage stays under 500MB, so a $5 box handles it with room to spare.
|
|
Pair it with the Telegram Gateway (covered in S09), and you can message Hermes from your phone anytime while it
|
|
responds from the VPS. The price of a coffee gets you a 24-hour AI assistant.
|
|
Serverless option: Hermes also supports Daytona and Modal as serverless backends. The environment hibernates
|
|
when idle and wakes automatically on incoming messages, driving inter-session costs toward zero. Great for light
|
|
usage when you still want to stay reachable. Set terminal: daytona or terminal: modal in config.yaml.
|
|
|
|
|
|
===== 第 33 页 =====
|
|
config.yaml Deep Dive
|
|
Regardless of how you install, all core configuration lives in one file: ~/.hermes/config.yaml .
|
|
A minimal working config looks like this:
|
|
# ~/.hermes/config.yaml
|
|
model:
|
|
provider: openrouter # Model provider
|
|
api_key: sk-or-xxxxx # Your API key
|
|
model: anthropic/claude-sonnet-4 # Model to use
|
|
terminal: local # Terminal backend (local/docker/ssh/daytona/modal)
|
|
gateway: # Messaging gateway (optional, details in §09)
|
|
telegram:
|
|
token: YOUR_BOT_TOKEN
|
|
discord:
|
|
token: YOUR_BOT_TOKEN
|
|
Not many fields — let's walk through each one.
|
|
provider and model
|
|
Hermes supports a wide range of model sources:
|
|
Provider
|
|
Recommended Models
|
|
Best For
|
|
OpenRouter
|
|
Claude Sonnet 4 / GPT-4o
|
|
200+ models available, flexible switching
|
|
Nous Portal
|
|
Hermes 3 series
|
|
Officially recommended, deeply integrated with the agent
|
|
OpenAI
|
|
GPT-4o / o3
|
|
Direct OpenAI API connection
|
|
z.ai / Zhipu
|
|
GLM-5
|
|
China-friendly option
|
|
Ollama
|
|
Hermes 3 8B/70B
|
|
Fully offline, privacy first
|
|
注意
|
|
Heads up: As of April 2026, Anthropic banned third-party tools from accessing Claude through Pro/Max
|
|
subscriptions. This affects Hermes, OpenClaw, and all other Agent frameworks. You can still use Claude via API
|
|
keys (pay-as-you-go), but it costs significantly more. Consider OpenRouter or Nous Portal's Hermes 3 series as
|
|
primary options.
|
|
My suggestion: start with OpenRouter so you can switch models freely and get a feel for them. Once you've settled
|
|
on your go-to model, connect directly to that provider's API to save the middleman fee.
|
|
terminal
|
|
|
|
|
|
===== 第 34 页 =====
|
|
Six backends determine where Hermes executes code:
|
|
local: Runs directly on your machine, simplest option
|
|
docker: Runs inside a container, isolated and secure
|
|
ssh: Connects to a remote server
|
|
daytona / modal: Serverless, spins up on demand
|
|
singularity: For HPC cluster environments
|
|
Most people should just go with local. If you're worried about the security of an agent executing code on your
|
|
machine, Docker is a solid middle ground.
|
|
|
|
|
|
===== 第 35 页 =====
|
|
Common Troubleshooting
|
|
注意
|
|
Install script hanging? Check your network connection. The script needs to download dependencies from GitHub
|
|
and PyPI. If you're in a region with restricted access, you may need to set up a proxy or use mirror sources.
|
|
注意
|
|
hermes command not found? The install script adds the command to your PATH, but if your shell config is
|
|
unusual (e.g., you use fish), you may need to manually run source ~/.bashrc or reopen your terminal.
|
|
核心建议
|
|
Want to verify the install? Run hermes --version — if you see a version number, you're set. The latest version is
|
|
v0.7.0.
|
|
注意
|
|
Docker container starts but nothing happens? Make sure ~/.hermes/config.yaml exists and has model info
|
|
configured. The container reads the config mapped in from the host. If config.yaml doesn't exist, Hermes will guide
|
|
you through creating one on startup.
|
|
核心建议
|
|
Security note for VPS deployments: If running on a VPS, consider setting terminal: docker so code executes
|
|
inside a container rather than directly on the host filesystem. A $5 VPS has plenty of power to run containers.
|
|
That's it for configuration. Hermes' design philosophy is to keep things minimal — one config.yaml handles
|
|
everything. No scattered environment variables, no layered config files. For an agent this feature-rich, keeping
|
|
config this simple is impressively restrained.
|
|
Next up, let's start talking. Hermes doesn't need you to configure everything before you can use it — install, add
|
|
your API key, launch, and you're ready for your first conversation.
|
|
|
|
|
|
===== 第 36 页 =====
|
|
§08 Your First Conversation: Letting Hermes Get to Know
|
|
You
|
|
First Conversation
|
|
Installed and launched — now what? This section walks you through the first conversation. The focus isn't
|
|
what you say, but what Hermes does behind the scenes.
|
|
Starting from a Blank Slate
|
|
Type hermes in your terminal and you'll see a clean chat interface. No onboarding flow, no setup wizard — just a
|
|
cursor waiting for you to talk.
|
|
Say anything:
|
|
Hey, I'm HuaShu. I run an AI-focused WeChat blog. I've been using Claude Code a lot and want to try o
|
|
Hermes will reply normally. But the interesting stuff is happening where you can't see it.
|
|
Memory Is Quietly Being Written
|
|
After the first exchange, take a look at the ~/.hermes/ directory:
|
|
~/.hermes/
|
|
├── config.yaml # Your configuration
|
|
├── state.db # SQLite database (conversation history + FTS5 index)
|
|
├── skills/ # Skills directory (still empty)
|
|
│ └── bundled/ # Built-in Skills
|
|
└── memories/ # Persistent memory (MEMORY.md + USER.md)
|
|
state.db already has content. That self-introduction you just typed, along with Hermes' reply, has been written
|
|
into the SQLite database with a FTS5 full-text index. Next time you start Hermes, it won't start from zero — it can
|
|
search and find this conversation.
|
|
This is different from ChatGPT's approach of "appearing to have memory but actually reloading all history every
|
|
time." Hermes retrieves on demand, only pulling up history when it's relevant. Even after months of
|
|
accumulated conversations, the database won't slow down.
|
|
Keep Talking, Trigger Deeper Memory
|
|
Chat for a few more rounds. Tell it about your work habits, for example:
|
|
|
|
|
|
===== 第 37 页 =====
|
|
I'm on macOS, my main editor is Cursor. I write articles in Markdown and prefer angle quotes over sta
|
|
Hermes will write this into its persistent memory layer. Mapping to the three-layer memory from S04: the
|
|
conversation content is session memory (what happened), while your preferences and habits are persistent
|
|
memory (who you are).
|
|
You don't need a special command to say "remember this." Hermes judges on its own which information is
|
|
worth persisting. If you've used Claude Code's auto-memory, the experience feels familiar. But Hermes is more
|
|
aggressive — it proactively strategizes what to remember.
|
|
|
|
|
|
===== 第 38 页 =====
|
|
Triggering the First Skill Creation
|
|
Give Hermes a slightly more complex task:
|
|
Convert this Markdown into WeChat-blog-compatible HTML, preserving bold and code block styling.
|
|
The first time, Hermes will figure it out as it goes. It might invoke the terminal to run a script, or generate the HTML
|
|
directly in the conversation.
|
|
After it's done, here's where it gets interesting. Check ~/.hermes/skills/ again:
|
|
~/.hermes/skills/
|
|
├── bundled/ # Built-in Skills
|
|
└── markdown-to-wechat.md # This is new!
|
|
Hermes automatically distilled the solution into a Skill. Open that markdown file and you'll see it records the
|
|
input format, conversion rules, and output requirements. Next time you ask for something similar, it'll invoke this
|
|
Skill directly instead of figuring it out from scratch.
|
|
This is the learning loop from S03 in action. You didn't teach it how to do things — it crystallized what it did into a
|
|
reusable capability on its own.
|
|
Skill self-improvement: If you're not happy with the result, tell it what needs adjusting. Hermes won't just fix the
|
|
current output — it'll update that Skill file too. Next time, the improved version runs automatically.
|
|
You Don't Need to Configure — Just Use It
|
|
Let's recap what just happened:
|
|
You chatted casually
|
|
→
|
|
Hermes built your profile
|
|
→
|
|
You gave it a task
|
|
→
|
|
Hermes auto-created a Skill
|
|
Through this entire process, you didn't write a single line of config, didn't edit a single file, didn't set a single
|
|
rule. That's a completely different experience from Claude Code asking you to hand-write CLAUDE.md, or
|
|
OpenClaw requiring you to configure yaml.
|
|
Of course, you absolutely can edit Skill files manually (covered in S10), but you really don't need to when starting
|
|
out. Just use it — Hermes will grow into the shape that fits you.
|
|
This is also what makes Hermes most distinctive. Other agent tools need you to figure out what you want, how to
|
|
configure it, and how to constrain it upfront. Hermes flips this around: you start using it first, and it forms its own
|
|
structure through the process of being used.
|
|
|
|
|
|
===== 第 39 页 =====
|
|
Next section — let's take Hermes beyond the terminal and onto your phone.
|
|
|
|
|
|
===== 第 40 页 =====
|
|
§09 Multi-Platform Access: Reach It Anywhere
|
|
Multi-Platform Access
|
|
Hermes doesn't just live in the terminal. Set up a Telegram Bot and reach it from your phone anytime. Add
|
|
Discord and Slack, and your team can use it too. The key: all platforms share the same brain.
|
|
Gateway: One Process, All Platforms
|
|
Hermes' multi-platform access is powered by the Messaging Gateway module. Instead of writing separate code for
|
|
each platform, it uses a single unified gateway process that listens to all configured platforms simultaneously.
|
|
The architecture looks like this:
|
|
Telegram
|
|
→
|
|
Messaging Gateway
|
|
←
|
|
Discord
|
|
Underneath the Gateway sits the same Hermes Agent instance, the same memory, the same set of Skills. A
|
|
message from Telegram and a command from the CLI are indistinguishable to Hermes.
|
|
Telegram Bot Setup (The Recommended Entry Point)
|
|
Why Telegram? It's the simplest to set up and the best mobile experience. Creating a Bot requires no approval
|
|
process — you get a Token instantly.
|
|
1
|
|
Create a Telegram Bot
|
|
Find @BotFather in Telegram, send /newbot , and follow the prompts to name it. BotFather will give you a
|
|
Token that looks like this:
|
|
123456789:ABCdefGhIJKlmNoPQRsTUVwxyz
|
|
2
|
|
Add It to config.yaml
|
|
# ~/.hermes/config.yaml
|
|
gateway:
|
|
telegram:
|
|
token: 123456789:ABCdefGhIJKlmNoPQRsTUVwxyz
|
|
|
|
|
|
===== 第 41 页 =====
|
|
3
|
|
Launch Hermes
|
|
hermes
|
|
Hermes automatically connects to the Telegram Gateway on startup. Send your Bot a message in Telegram
|
|
and it'll respond.
|
|
Three steps, under two minutes total. If Hermes is running on a VPS (the $5 setup from S07), this gives you a 24-
|
|
hour online, always-reachable, persistent-memory personal AI assistant. $5/month plus model API costs.
|
|
核心建议
|
|
Telegram also supports voice messages. Send a voice memo and Hermes will automatically transcribe it to text
|
|
before processing. Think of something during your commute? Just say it — no typing needed.
|
|
|
|
|
|
===== 第 42 页 =====
|
|
Discord and Slack Integration
|
|
The process is similar to Telegram, with differences mainly in how you obtain the Token.
|
|
Discord
|
|
Go to the Discord Developer Portal, create an Application, and grab the Token from the Bot page:
|
|
gateway:
|
|
discord:
|
|
token: YOUR_DISCORD_BOT_TOKEN
|
|
Invite the Bot to your Server and it can respond in channels. Great for team collaboration — have Hermes help
|
|
review code in the dev channel, or summarize data in the operations channel.
|
|
Slack
|
|
gateway:
|
|
slack:
|
|
token: xoxb-YOUR-SLACK-BOT-TOKEN
|
|
You'll need to create an App in the Slack App management page and install it to your Workspace. Permission setup
|
|
is a bit more involved than Telegram, but it's the more enterprise-appropriate option that IT departments are
|
|
comfortable with.
|
|
More platforms available: Hermes also supports WhatsApp, Signal, Email, SMS (Twilio), Home Assistant,
|
|
Mattermost, Matrix, DingTalk, Feishu/Lark, WeCom, and Open WebUI — 14 platforms in total. Full list on the official
|
|
docs' Messaging page. Most just need the corresponding Token added to config.yaml.
|
|
Cross-Platform Conversation Continuity
|
|
This is the most practical feature of Hermes' multi-platform design.
|
|
Say you're on your morning commute and message Hermes via Telegram:
|
|
Research Hermes Agent deployment options and put together a doc for me.
|
|
Hermes starts working, storing the research results in memory. You get to the office and open your terminal:
|
|
How's that research coming? Show me what you've got.
|
|
Hermes knows exactly what you're talking about. It doesn't distinguish which platform a message came from —
|
|
all platforms share the same Agent instance and the same memory. What you said on Telegram can be continued
|
|
in the CLI. What was discussed in a Discord channel can be referenced from Slack.
|
|
|
|
|
|
===== 第 43 页 =====
|
|
This is nothing like having different ChatGPT windows on different devices where you have to re-explain the
|
|
context every time. Hermes has one brain, no matter which door you walk through.
|
|
|
|
|
|
===== 第 44 页 =====
|
|
A Practical Deployment Setup
|
|
Pulling together the previous sections, a typical deployment looks like this:
|
|
$5 VPS (Ubuntu 22.04)
|
|
├── Hermes Agent Core
|
|
├── Messaging Gateway
|
|
│ ├── Telegram Bot (reachable from your phone)
|
|
│ ├── Discord Bot (team collaboration)
|
|
│ └── Slack App (enterprise use)
|
|
├── ~/.hermes/
|
|
│ ├── state.db (all conversation history)
|
|
│ ├── skills/ (capabilities that accumulate automatically)
|
|
│ └── config.yaml (one file, everything configured)
|
|
└── Model calls → OpenRouter API
|
|
Total cost: VPS $5/month + model API fees (light usage runs $2-5/month). Less than a fancy cup of coffee per
|
|
month for an AI assistant with memory, real capabilities, and 24/7 availability.
|
|
核心建议
|
|
On model API costs: If you want to cut costs further, you can run open-source models on the VPS via Ollama (like
|
|
Hermes 3 13B). A $5 VPS might not have enough RAM for large models, but a $10-15/month VPS can handle 13B,
|
|
and after that model calls are completely free.
|
|
Automated Scheduling
|
|
Beyond passively responding to messages, the Gateway also supports cron scheduling. You can have Hermes
|
|
automatically send a news briefing every morning at 8am, or automatically summarize the week's GitHub commits
|
|
every Friday afternoon.
|
|
Scheduled task results get pushed through the Gateway to whatever platform you specify. It doesn't just wait for
|
|
you to talk — it can work proactively.
|
|
At this point, you have a working Hermes setup: installed locally or on a VPS, model configured, reachable from
|
|
your phone. Next up: how to manually create and optimize Skills (S10), and how to integrate more tools via MCP
|
|
(S11), to make this assistant even more powerful.
|
|
|
|
|
|
===== 第 45 页 =====
|
|
§10 Custom Skills: Teaching Hermes New Tricks
|
|
Custom Skills: Teaching Hermes New Tricks
|
|
Hermes's learning loop creates Skills automatically, but you can also teach it manually. This section covers
|
|
how to write Skills, install them from the community, and port your Claude Code Skills over.
|
|
What Exactly Is a Skill
|
|
In Hermes, a Skill is just a markdown file. No framework to learn, no API to call -- just text that tells Hermes what
|
|
to do in a specific scenario.
|
|
The idea is the same as CLAUDE.md: define behavior in natural language. The difference is that CLAUDE.md
|
|
applies globally, while Skills activate on demand. Ask Hermes to write your weekly report, and it loads the "weekly
|
|
report" Skill; ask it to do a code review, and it loads a different one.
|
|
Each Skill lives in its own folder under ~/.hermes/skills/ , with a SKILL.md file as the entry point. A Skill folder
|
|
can also contain references/ , templates/ , and scripts/ subdirectories for supporting files.
|
|
Creating a Skill by Hand
|
|
Let's write a Skill that's actually useful: making Hermes enforce a consistent Git commit message format.
|
|
Create a folder called git-commit-style/ in ~/.hermes/skills/ , and inside it create a SKILL.md file:
|
|
|
|
|
|
===== 第 46 页 =====
|
|
---
|
|
name: git-commit-style
|
|
description: Enforce a consistent Git commit message format
|
|
version: "1.0.0"
|
|
---
|
|
# Git Commit Style
|
|
## Trigger
|
|
Activate when the user asks me to commit code, write a commit message, or review commit history.
|
|
## Rules
|
|
### Commit Message Format
|
|
- First line: type(scope): summary (50 chars max)
|
|
- Blank line
|
|
- Body: explain WHY, not WHAT
|
|
### Type Enum
|
|
- feat: new feature
|
|
- fix: bug fix
|
|
- refactor: restructure (no behavior change)
|
|
- docs: documentation
|
|
- test: tests
|
|
- chore: build/toolchain
|
|
### Constraints
|
|
- Body in plain language, types in English
|
|
- Don't write "modified XX file" -- that's noise
|
|
- One commit, one thing
|
|
## Example
|
|
```
|
|
feat(auth): add WeChat QR code login
|
|
Previously users could only log in with a phone number, which meant WeChat users
|
|
had to bind a phone first. Now they scan a QR code and they're in -- unbound users
|
|
get an account created automatically.
|
|
```
|
|
Save it, and the next time you say "commit these changes," Hermes follows this format. No extra setup needed.
|
|
核心建议
|
|
Skill triggering is automatic. You don't need to say "use XX Skill" -- Hermes matches the most relevant Skill to
|
|
your request on its own. Under the hood, it uses FTS5 full-text search plus semantic understanding.
|
|
Anatomy of a Good Skill
|
|
After writing dozens of Skills, I've found they share a common structure:
|
|
|
|
|
|
===== 第 47 页 =====
|
|
Section
|
|
Purpose
|
|
Required?
|
|
Title
|
|
Lets Hermes quickly identify what the Skill does
|
|
Yes
|
|
Trigger
|
|
When to activate this Skill
|
|
Strongly recommended
|
|
Rules
|
|
Concrete steps, constraints, formats
|
|
Yes
|
|
Example
|
|
A complete input-to-output example
|
|
Strongly recommended
|
|
Don'ts
|
|
Explicit boundaries to prevent drift
|
|
Optional
|
|
The more specific your trigger, the better the hit rate. "When the user mentions code" is too vague; "When the
|
|
user asks me to commit code, write a commit message, or review commit history" is much better.
|
|
Installing Community Skills from the Skills Hub
|
|
Hermes has a built-in Skills Hub where community developers share ready-made Skills. Installing them is
|
|
straightforward:
|
|
1
|
|
Browse Available Skills
|
|
Just ask Hermes in conversation: "What community Skills are available?" It'll list popular Skills from the Hub,
|
|
grouped by category. You can also narrow it down: "Any Python-related Skills?"
|
|
2
|
|
Install
|
|
Found one you like? Tell Hermes "Install XX Skill." It downloads the Skill folder to ~/.hermes/skills/ and it
|
|
takes effect immediately. No restart needed.
|
|
3
|
|
Customize
|
|
An installed Skill is just a folder with a SKILL.md file -- open it and edit away. Community Skills are a starting
|
|
point; making them yours is the goal.
|
|
Hermes ships with 40+ bundled Skills covering MLOps, GitHub workflows, research assistance, and more. They're
|
|
built in -- no extra installation required.
|
|
Debugging Skills
|
|
You wrote a Skill -- how do you know it's actually firing?
|
|
Just ask. Say "What Skills do you have loaded right now?" and Hermes tells you which Skills are active. If the one
|
|
you expected isn't there, your trigger is probably too narrow.
|
|
Check the logs. Log files record which Skills matched each request, why they were selected, and why others were
|
|
skipped. Logs live in ~/.hermes/logs/ .
|
|
|
|
|
|
===== 第 48 页 =====
|
|
Test incrementally. Don't jump straight to edge cases. Start with the simplest request to confirm the trigger and
|
|
basic behavior, then try boundary conditions.
|
|
注意
|
|
Skills can conflict. If two Skills have overlapping triggers, Hermes picks the one with the higher match score -- but
|
|
the result may not be what you expect. If behavior seems off, check for Skill conflicts first.
|
|
Hands-On: Porting a Claude Code Skill to Hermes
|
|
Here's a very practical need: you've built up a collection of great Skills in Claude Code and don't want to start from
|
|
scratch after switching to Hermes. The agentskills.io standard makes Skills portable across Agents, so migration
|
|
cost is low.
|
|
Let me walk through a real example. I had a "WeChat blog proofreading" Skill in Claude Code -- its core logic runs
|
|
three proofreading passes (facts, style, details). The SKILL.md looked roughly like this:
|
|
---
|
|
name: proofreading
|
|
description: Three-pass proofreading for articles
|
|
version: "1.0.0"
|
|
---
|
|
# Article Proofreading
|
|
## Trigger
|
|
Activate when the user mentions "proofread," "reduce AI tone," "too AI-like," or "polish."
|
|
## Proofreading Flow
|
|
### Pass 1: Fact Check
|
|
- Verify all data, dates, product names
|
|
- Flag anything uncertain
|
|
### Pass 2: Style Check
|
|
- Remove AI-favorite phrases (firstly/secondly/in conclusion)
|
|
- Break up AI sentence patterns
|
|
- Replace formal vocabulary with conversational language
|
|
### Pass 3: Detail Polish
|
|
- Keep sentences to 15-25 words
|
|
- Keep paragraphs to 3-5 lines on a phone screen
|
|
- Bold ~10 key sentences for scanability
|
|
Copy this file to ~/.hermes/skills/proofreading/SKILL.md , and Hermes can use it right away. No format changes,
|
|
no API adapters -- everyone uses markdown with the same semantic structure.
|
|
If your Skill references tools specific to Claude Code (like a particular MCP Server), you'll need to swap those for
|
|
Hermes equivalents. But the core logic, triggers, and rules are all portable.
|
|
|
|
|
|
===== 第 49 页 =====
|
|
Why agentskills.io matters: Skills are no longer locked to a single Agent. Skills you built in Claude Code, Cursor, or
|
|
Gemini CLI work in Hermes. And vice versa. You can switch Agents without worrying about migration cost.
|
|
Next up: MCP integration. Skills teach Hermes how to do things; MCP lets it connect to external tools. Put them
|
|
together, and the range of scenarios you can cover gets very wide.
|
|
|
|
|
|
===== 第 50 页 =====
|
|
§11 MCP Integration: Connecting Your Tool Stack
|
|
MCP Integration: Connecting Your Tool Stack
|
|
Hermes's 40+ built-in tools are already quite capable, but real-world workflows go far beyond that. MCP lets
|
|
Hermes plug into GitHub, databases, Slack, Jira, and thousands of other external services -- without writing
|
|
a single line of adapter code.
|
|
What MCP Is and Why It Matters
|
|
MCP stands for Model Context Protocol, an open standard proposed by Anthropic in late 2024. Think of it as the
|
|
USB port of the AI tool world: as long as an MCP Server implements this protocol, any MCP-compatible Agent can
|
|
call the tools it provides.
|
|
For Hermes, MCP means no need to build a custom tool for every external service. Want to connect to GitHub?
|
|
Install a GitHub MCP Server. Need to query a database? Install a database MCP Server. The community already has
|
|
thousands of ready-made Servers.
|
|
Two Connection Modes: stdio and HTTP
|
|
Hermes supports two ways to connect to MCP Servers, depending on where the Server runs.
|
|
Mode
|
|
Server Location
|
|
Best For
|
|
Performance
|
|
stdio
|
|
Local subprocess
|
|
Local tools, file system, databases
|
|
Fast, no network overhead
|
|
HTTP (StreamableHTTP)
|
|
Remote server
|
|
Cloud services, shared team Servers
|
|
Depends on network
|
|
stdio is enough for most cases. The MCP Server runs as a Hermes subprocess, communicating over stdin/stdout --
|
|
fast and simple to set up. HTTP is for when the Server needs to be independently deployed or shared across
|
|
multiple Agents.
|
|
Configuration goes in the mcp_servers section of config.yaml :
|
|
|
|
|
|
===== 第 51 页 =====
|
|
# stdio mode
|
|
mcp_servers:
|
|
github:
|
|
command: "npx"
|
|
args: ["-y", "@modelcontextprotocol/server-github"]
|
|
env:
|
|
GITHUB_PERSONAL_ACCESS_TOKEN: "ghp_xxxxx"
|
|
# HTTP mode
|
|
mcp_servers:
|
|
remote-tools:
|
|
url: "https://your-server.com/mcp"
|
|
transport: "streamable-http"
|
|
Hands-On: Connecting GitHub MCP
|
|
GitHub is one of the most common MCP integrations. Once connected, Hermes can create Issues, open PRs, review
|
|
code, and manage project boards directly.
|
|
1
|
|
Create a GitHub Token
|
|
Go to GitHub Settings -> Developer settings -> Personal access tokens and generate a token. At minimum,
|
|
check repo and read:org . If you want to manage Issues and PRs, enable write access for issues and
|
|
pull_requests too.
|
|
2
|
|
Configure config.yaml
|
|
Add the GitHub MCP Server to config.yaml :
|
|
mcp_servers:
|
|
github:
|
|
command: "npx"
|
|
args: ["-y", "@modelcontextprotocol/server-github"]
|
|
env:
|
|
GITHUB_PERSONAL_ACCESS_TOKEN: "ghp_your_token_here"
|
|
Pro tip: store the token in an environment variable instead of hardcoding it. You can reference it with
|
|
${GITHUB_TOKEN} .
|
|
3
|
|
Restart Hermes and Verify
|
|
After restarting, ask Hermes "List my GitHub repos" or "Show recent Issues in the XX repo." If it returns the
|
|
right info, you're connected.
|
|
|
|
|
|
===== 第 52 页 =====
|
|
4
|
|
Daily Use
|
|
Once connected, you can operate GitHub with natural language. For example:
|
|
"Create an Issue in the alchaincyf/my-app repo titled 'Fix login page redirect bug'"
|
|
"Look at this PR's changes and do a code review"
|
|
"What new Issues were opened this past week? Group them by label"
|
|
The GitHub MCP provides tools for repo management, Issue operations, PR reviews, code search, branch
|
|
management, and more. You don't need to remember tool names -- just describe what you need in plain
|
|
language and Hermes picks the right tool.
|
|
Hands-On: Connecting a Database
|
|
Databases are another high-frequency use case. Once connected, Hermes can query data, generate reports, and
|
|
analyze trends -- no hand-written SQL required.
|
|
Using PostgreSQL as an example:
|
|
mcp_servers:
|
|
postgres:
|
|
command: "npx"
|
|
args: ["-y", "@modelcontextprotocol/server-postgres"]
|
|
env:
|
|
POSTGRES_CONNECTION_STRING: "postgresql://user:pass@localhost:5432/mydb"
|
|
Once configured, you can ask Hermes "How many users signed up this month?" or "Show me the daily order
|
|
revenue trend for the last 30 days" -- it generates SQL, runs it, and returns the results.
|
|
注意
|
|
Database MCP has read-write access by default. If you only want Hermes to query without modifying data, connect
|
|
with a read-only database account. This is especially important for production databases.
|
|
SQLite and MySQL have their own MCP Servers too. Configuration is nearly identical -- just swap the Server
|
|
package name and connection string.
|
|
Per-Server Tool Filtering
|
|
Once you connect multiple MCP Servers, the available tool list grows fast. A single GitHub Server exposes a dozen-
|
|
plus tools; add databases, file systems, and Slack, and you could have fifty to a hundred.
|
|
Too many tools degrades Agent decision quality. Matching accuracy with 100 tools is lower than with 20. And
|
|
some tools you don't want the Agent touching at all.
|
|
Hermes supports per-server tool filtering -- specify in the config which tools each Server should expose:
|
|
|
|
|
|
===== 第 53 页 =====
|
|
mcp_servers:
|
|
github:
|
|
command: "npx"
|
|
args: ["-y", "@modelcontextprotocol/server-github"]
|
|
env:
|
|
GITHUB_PERSONAL_ACCESS_TOKEN: "ghp_xxxxx"
|
|
allowed_tools:
|
|
- "list_issues"
|
|
- "create_issue"
|
|
- "get_pull_request"
|
|
- "create_pull_request_review"
|
|
This way, even though the GitHub MCP Server offers high-privilege tools like deleting repos or changing settings,
|
|
Hermes can't use them. The principle of least privilege matters more in the Agent era than ever before.
|
|
When to Use MCP vs. Native Tools
|
|
Hermes has 40+ built-in tools, and MCP opens up thousands more. How do you choose?
|
|
推荐
|
|
Use native tools for:
|
|
Terminal commands, file operations, web search, image
|
|
generation, memory management, sub-Agent
|
|
delegation. These built-in tools are deeply optimized and
|
|
tightly integrated with the learning loop and memory
|
|
system -- fast response, predictable behavior.
|
|
不推荐
|
|
Use MCP for:
|
|
GitHub, databases, Slack, Jira, Google Drive, and other
|
|
external services. These require specific API protocols,
|
|
making MCP the right choice. You could use the terminal
|
|
to run CLIs instead, but MCP provides structured
|
|
input/output, which means higher accuracy.
|
|
A simple rule of thumb: if Hermes already has the capability built in, use built-in; if you need to interact with an
|
|
external service, use MCP.
|
|
Some scenarios work either way -- like Git operations. The terminal tool can run git commands directly, and the
|
|
GitHub MCP can operate on repos too. The difference: terminal works locally, great for everyday commits and
|
|
pushes in your current repo; GitHub MCP works through the API, better for cross-repo management, batch Issue
|
|
operations, and PR reviews that need platform-level capabilities.
|
|
Practical advice: Don't connect a dozen MCP Servers on day one. Start with one or two you use most (GitHub,
|
|
database), get comfortable, then add more. Every additional Server expands the tool selection space and
|
|
lengthens the decision path.
|
|
MCP + Skills: The Combo
|
|
MCP solves "what can I connect to," Skills solve "how to use it." They work better together.
|
|
Example: you connect the GitHub MCP and create a "Code Review" Skill. The Skill defines review criteria (naming
|
|
conventions, error handling, test coverage), while MCP provides the ability to read PR diffs. Combined, Hermes can
|
|
automatically review code against your standards.
|
|
|
|
|
|
===== 第 54 页 =====
|
|
Another example: the database MCP lets Hermes run SQL, and a "Weekly Report" Skill defines the report format
|
|
and key metrics. Pair that with a Friday afternoon cron job, and Hermes automatically queries the data, generates
|
|
the report, and posts it to Slack. MCP, Skills, and native tools working together -- that's where the real power is.
|
|
At this point, you know all of Hermes's core capabilities. The next Part gets into real-world scenarios, showing what
|
|
these capabilities look like when combined.
|
|
|
|
|
|
===== 第 55 页 =====
|
|
§12 Personal Knowledge Assistant: The Power of Cross-
|
|
Session Memory
|
|
Personal Knowledge Assistant: The Power of Cross-Session Memory
|
|
How big is the gap between an AI assistant that actually remembers things and one you have to re-introduce
|
|
yourself to every day? Let's find out with a project that spans three weeks.
|
|
Three Weeks of Research, Starting from Scratch Every Time
|
|
Say you're researching a new domain. You're an indie developer trying to figure out AI Agent deployment options
|
|
for 2026: local, cloud, Serverless -- what are the gotchas for each.
|
|
Week one, you ask three or four questions, covering Docker deployment memory usage, VPS pricing, Daytona's
|
|
free tier limitations. This information ends up scattered across different conversations.
|
|
Week two, you want to dig deeper into Serverless. You open ChatGPT or Claude, and what's the first thing you do?
|
|
Re-explain what you're working on.
|
|
"I'm researching AI Agent deployment options. Last week I looked at Docker and VPS, now I want to understand
|
|
Serverless. I found that Daytona has a free tier but with limitations..."
|
|
Every new conversation costs 3-5 minutes of context-setting. This isn't an AI capability problem -- it's an
|
|
architecture problem: traditional AI has no cross-session memory, so every conversation starts from a blank slate.
|
|
What Hermes Remembers
|
|
Same scenario, but with Hermes.
|
|
After week one's conversations, Hermes's three memory layers have recorded different things:
|
|
Memory Layer
|
|
What It Records
|
|
Purpose
|
|
Session memory (SQLite
|
|
+ FTS5)
|
|
What you asked, what it looked up, raw conversation
|
|
text
|
|
Precise retrieval when details are
|
|
needed
|
|
Persistent memory
|
|
"User is researching AI Agent deployment, ruled out
|
|
option X, prefers low cost"
|
|
Auto-loads context for the next
|
|
conversation
|
|
Skill memory
|
|
"Research tasks: list dimensions first -> dig into each ->
|
|
summarize per round"
|
|
Methodology reuse
|
|
Week two, you open Hermes and just say "Let's continue with the Serverless options." No need to re-explain
|
|
anything -- persistent memory already knows what you're doing. It might even proactively remind you: "Last
|
|
|
|
|
|
===== 第 56 页 =====
|
|
week you mentioned Daytona's free tier had limitations -- want to check if the policy has been updated?"
|
|
This isn't magic; it's FTS5 full-text search at work. Hermes doesn't stuff all of last week's conversations into the
|
|
context -- that would waste tokens. Instead, it retrieves the most relevant historical snippets based on your current
|
|
question.
|
|
Retrieval vs. Full-Context Loading
|
|
This design choice is worth unpacking.
|
|
Many people assume "memory" means cramming all conversation history into the prompt. Claude Code's auto-
|
|
memory does work that way: key info gets written into MEMORY.md, which is fully loaded at startup. Fine for
|
|
coding tasks.
|
|
But the knowledge assistant use case is different. Three weeks of research conversations could be tens of
|
|
thousands of words. Loading everything causes two problems: token costs explode, and information overload
|
|
actually degrades answer quality. Large models have uneven attention distribution across very long contexts,
|
|
and key information gets drowned out.
|
|
Hermes's approach: persistent memory stores summaries (a few hundred words), and when details are
|
|
needed, FTS5 searches the raw conversations, injecting only the most relevant snippets into context. It's like
|
|
carrying a one-page cheat sheet and going back to the filing cabinet only when you need specifics.
|
|
推荐
|
|
Hermes approach: Persistent memory (summaries) +
|
|
on-demand retrieval (FTS5). Token consumption stays
|
|
controlled, information stays precise.
|
|
不推荐
|
|
Full-context approach: Stuff all history into the prompt.
|
|
Works short-term, but after three weeks the prompt is
|
|
maxed out and costs double.
|
|
Honcho: It Knows You Better Than You Know Yourself
|
|
If you enable Honcho user modeling, the depth of memory goes up another level.
|
|
Honcho doesn't just record what you said -- it infers things you didn't say. For instance, if you consistently pick
|
|
the cheapest option across three research sessions, Honcho infers "this user is cost-sensitive." Next time it
|
|
recommends something, pricing info gets surfaced first.
|
|
This dialectical modeling continuously derives insights about you -- from technical skill level and preference
|
|
patterns to communication habits -- building a representation that deepens over time. The longer you use it, the
|
|
more accurately it understands you.
|
|
HuaShu's experience: After two weeks, Hermes started automatically giving me shorter, punchier replies --
|
|
because it noticed I tend to want conclusions rather than lengthy analysis. This adaptation is gradual; you don't
|
|
have to explicitly configure anything.
|
|
|
|
|
|
===== 第 57 页 =====
|
|
The Experience Gap vs. Traditional AI
|
|
Here's an analogy to sum it up: Traditional AI is like a hotel front desk -- different person every day, and you re-
|
|
introduce yourself each time. Hermes is like your personal assistant who's known you for three months, knows
|
|
you drink black coffee, hates writing weekly reports, and what project you're researching right now.
|
|
Specifically for the knowledge assistant use case, the gap shows up in three ways:
|
|
1
|
|
Zero Startup Cost
|
|
No context-setting every time. Say "continue" and you continue.
|
|
2
|
|
Research Has Continuity
|
|
Options you ruled out last week don't get re-recommended. Information you already confirmed doesn't get
|
|
re-verified. The whole research effort feels like a single thread instead of a pile of disconnected dots.
|
|
3
|
|
Methodology Compounds
|
|
The approach Hermes develops while helping you with your first research project (list dimensions, dig into
|
|
each, summarize per round) gets saved as a Skill. Your second research project reuses it automatically --
|
|
no re-teaching needed.
|
|
核心建议
|
|
Cross-session memory isn't a silver bullet. If the task is one-off -- like "translate this paragraph" -- Hermes's
|
|
memory system offers zero advantage. The value of memory scales with task duration. The longer and more
|
|
context-heavy the task, the bigger the efficiency gain from memory.
|
|
|
|
|
|
===== 第 58 页 =====
|
|
§13 Dev Automation: From Code Review to Deployment
|
|
Dev Automation: From Code Review to Deployment
|
|
Claude Code writes the code, Hermes watches the pipeline. These two tools aren't either-or -- they each own
|
|
a different part of the workflow.
|
|
A Developer's Morning
|
|
9 AM, you open your laptop. Three Telegram messages pop up. Not from a coworker -- from Hermes:
|
|
"Last night at 11:17 PM, a PR was merged into main -- 387 new lines of code. I reviewed it and found two issues: the
|
|
auth module's token expiration logic doesn't handle an edge case, and test coverage dropped from 82% to 76%.
|
|
Full report saved to the project Skill."
|
|
"At 2:40 AM, the CI pipeline ran a regression test suite. 3 cases failed. 2 were introduced by yesterday's PR, 1 is a
|
|
known flaky test."
|
|
"Today's daily standup notes are drafted, based on yesterday's 4 commits and 2 PRs. Needs your sign-off before
|
|
sending."
|
|
This isn't hypothetical. Hermes's cron scheduling + GitHub MCP + memory system lets it keep working while
|
|
you sleep.
|
|
Automated Code Review
|
|
Traditional code review has two pain points: latency -- you open a PR and wait for a reviewer to find time; and
|
|
inconsistency -- different reviewers focus on different things, one checks architecture while another checks
|
|
formatting.
|
|
Setting up automated code review with Hermes is straightforward:
|
|
1
|
|
Connect GitHub MCP
|
|
Configure the GitHub MCP Server in config.yaml, and Hermes can read your repo's PRs, diffs, and issues.
|
|
2
|
|
Set Up cron Scheduling
|
|
Tell Hermes in plain language: "Check main branch for new PRs every 6 hours and do a code review." It
|
|
creates the cron job automatically.
|
|
3
|
|
Define Review Standards
|
|
Write your coding standards as a Skill. For example: "Functions must be under 50 lines," "Error handling
|
|
must use custom types," "All API endpoints must have tests." Hermes reviews every PR against these
|
|
standards.
|
|
|
|
|
|
===== 第 59 页 =====
|
|
Step three is the key. Review standards are a Skill, and they evolve automatically based on your feedback. You
|
|
flag an issue Hermes missed, and next time it watches for that pattern. Traditional lint rules are static; Hermes's
|
|
review standards are alive.
|
|
Test Generation and Execution
|
|
There's a fundamental difference between how Hermes does testing and how Claude Code does it: with Claude
|
|
Code, you say "write tests for this function," it writes them, and you verify. Hermes discovers which functions
|
|
lack tests on its own, writes the tests, runs them, fixes failures, and delivers a report.
|
|
This capability comes from combining several tools:
|
|
Tool
|
|
Role
|
|
file
|
|
Scans the codebase to find modules without tests
|
|
code_execution
|
|
Runs tests in a sandbox
|
|
terminal
|
|
Generates coverage reports
|
|
memory
|
|
Remembers which modules have been reviewed and which tests are flaky
|
|
Pair it with cron for weekly coverage checks. Every Monday morning it runs an automatic scan, and if coverage
|
|
drops below the threshold, it sends a notification.
|
|
Automated Daily and Weekly Reports
|
|
This feature sounds simple, but it's addictive once you start using it.
|
|
Hermes pulls the day's commit history, PR statuses, and issue changes through GitHub MCP, combines that with
|
|
conversation memory from your discussions, and generates a daily report so accurate you can approve it
|
|
without even thinking back through your day.
|
|
Weekly reports are even more interesting. Thanks to cross-session memory, Hermes can see the full arc of the
|
|
week: what bug you fixed Monday, why you changed the architecture plan Wednesday, why Friday's PR got
|
|
rejected. It's not just a list of commits -- it's a narrative with cause and effect.
|
|
Pro tip: When Hermes generates the daily report, have it also update a project Skill that logs the week's technical
|
|
decisions. When it writes the weekly report, it can reference those decisions directly instead of guessing what you
|
|
were thinking from commit messages.
|
|
|
|
|
|
===== 第 60 页 =====
|
|
How Claude Code and Hermes Divide the Work
|
|
These two tools aren't competing. They're good at completely different things:
|
|
Dimension
|
|
Claude Code
|
|
Hermes Agent
|
|
Interaction mode
|
|
You're right there, real-time conversation
|
|
Runs in the background, reports on schedule
|
|
Strengths
|
|
Writing code, refactoring, debugging
|
|
Monitoring, auditing, summarizing, scheduling
|
|
Time horizon
|
|
Completed within a single session
|
|
Runs continuously across days and weeks
|
|
Trigger
|
|
You initiate it
|
|
cron or event-driven
|
|
In one sentence: Claude Code is the craftsman, Hermes is the butler. The craftsman builds things; the butler
|
|
makes sure everything stays on track. You wouldn't ask the butler to lay bricks, and you wouldn't ask the
|
|
craftsman to do the night watch.
|
|
Claude Code writes code and opens a PR
|
|
→
|
|
Hermes auto-reviews the PR
|
|
→
|
|
Hermes runs tests to verify
|
|
→
|
|
Hermes generates the daily report
|
|
Once this pipeline is running, your focus shifts from "write code + review code + run tests + write reports" to "write
|
|
code + confirm results." Everything in between is automated.
|
|
注意
|
|
Hermes's code review is a supplement, not a replacement. It's great at catching pattern-based issues
|
|
(inconsistent naming, missing tests, overlooked edge cases), but architectural decisions still need a human eye.
|
|
Don't skip manual review just because you have automation -- at minimum, keep a human in the loop for core
|
|
modules.
|
|
|
|
|
|
===== 第 61 页 =====
|
|
§14 Content Creation: From Research to Final Draft
|
|
Content Creation: From Research to Final Draft
|
|
Using AI to write articles isn't new. Having AI remember your writing style, research habits, even reader
|
|
feedback, and continuously evolve into your personal editor — that's how Hermes does content.
|
|
The Life Cycle of an Article
|
|
I've written over 100 WeChat blog articles with Claude Code. The workflow is well-established: give it a topic, it
|
|
searches, drafts, proofreads, and generates images. The whole process takes about 2 to 3 hours.
|
|
But one problem never went away: every time I started a new session, it forgot the lessons from the last
|
|
article.
|
|
Last week, I told it to stop using a certain cliched phrase. It complied. This week, writing a new article, the same
|
|
phrase was back. I added a banned-words list to CLAUDE.md, which helped — but CLAUDE.md is a static file. It
|
|
doesn't update itself.
|
|
This is exactly where Hermes does content creation differently.
|
|
Ongoing Content Projects
|
|
Say you're working on a content series — five consecutive articles about AI Agents. The traditional way, each
|
|
article is independent: you re-explain the reader persona, re-define the style, re-state which topics were already
|
|
covered.
|
|
With Hermes, the series works completely differently:
|
|
After the first article is done, Hermes's memory system records several things: the series positioning, target
|
|
audience, your editing preferences (like how you broke all its long sentences into short ones), and which technical
|
|
concepts were already explained in article one.
|
|
When you start the second article, you just say "write the next one in this series, topic is XXX." It knows what style
|
|
to use, which concepts need no re-explanation, and what you were unhappy about last time.
|
|
By the fifth article, its understanding of your writing preferences is remarkably precise. Not because you configured
|
|
anything — because it learned from your feedback on its own.
|
|
Parallel Research with Sub-Agents
|
|
For a deep technical article, research often takes more time than writing. The traditional approach is linear: search
|
|
topic A, organize it, search topic B, organize that, then topic C.
|
|
Hermes's delegate_task tool makes this parallel. Say you're writing an AI Agent comparison piece — you can
|
|
dispatch three sub-agents simultaneously:
|
|
|
|
|
|
===== 第 62 页 =====
|
|
Sub-Agent 1
|
|
Research Hermes architecture
|
|
→
|
|
Sub-Agent 2
|
|
Research OpenClaw ecosystem
|
|
→
|
|
Sub-Agent 3
|
|
Collect community feedback
|
|
Three sub-agents working simultaneously, each returning their research results. The main agent consolidates
|
|
everything into structured research material. Research that used to take over an hour? Done in 20 minutes.
|
|
Each sub-agent can be assigned a different toolset. The one researching architecture needs web search and
|
|
browser; the one collecting community feedback might only need web search. Restricted toolsets aren't just a
|
|
safety design — they're an efficiency design — sub-agents don't get distracted by having too many tools.
|
|
Skills That Accumulate Writing Style
|
|
This is the most valuable part of how Hermes handles content creation.
|
|
Traditional AI writing tools control style through prompts. You write in the prompt: "conversational tone,
|
|
sentences under 20 words, avoid AI-typical phrases." You have to write this every time, or maintain a very long
|
|
system prompt.
|
|
Hermes stores style rules as a Skill. Initially, this Skill might have just a few simple rules: don't use cliched
|
|
summary phrases, keep paragraphs to 3-5 lines, use casual words like "I think" or "actually."
|
|
Here's the key: this Skill self-improves.
|
|
When you proofread a draft and make edits, Hermes observes and learns. If you change a stiff phrasing to
|
|
something more natural three times in a row, it adds a rule to the Skill: "avoid formal verb constructions." If you
|
|
delete a forced inspirational ending it wrote, it learns "don't force an uplifting conclusion."
|
|
A month later, this writing style Skill has accumulated dozens of rules, all from your real feedback. It becomes
|
|
your personal editorial handbook — and it maintains itself automatically.
|
|
核心建议
|
|
Hermes's Skill self-improvement isn't a black box. Every Skill update is visible as a diff in the ~/.hermes/skills/
|
|
directory. If a rule drifts in the wrong direction, you can manually correct it, and Hermes incorporates your
|
|
correction into its learning.
|
|
|
|
|
|
===== 第 63 页 =====
|
|
How Is This Different from Writing with Claude Code
|
|
I'm using both tools for content creation simultaneously. The difference is clear:
|
|
Dimension
|
|
Claude Code
|
|
Hermes Agent
|
|
Best for
|
|
Standalone articles, one-off tasks
|
|
Content series, ongoing projects
|
|
Style control
|
|
CLAUDE.md + manual maintenance
|
|
Skills that auto-accumulate and evolve
|
|
Research efficiency
|
|
Linear search
|
|
Parallel research via sub-agents
|
|
Context continuity
|
|
Relies on auto-memory, limited capacity
|
|
Three-layer memory, on-demand retrieval
|
|
Learning ability
|
|
Doesn't learn; rules are manually written
|
|
Learns automatically from your feedback
|
|
This isn't to say Claude Code is worse. For a single article, Claude Code's interactive experience is smoother — you
|
|
see edits in real time and give feedback on the fly. Hermes's advantage is in the long game. Write two articles a
|
|
week for three months, and by article ten Hermes is dramatically better than article one. Claude Code's article ten
|
|
performs about the same as article one.
|
|
HuaShu's approach: I use Claude Code for one-off sponsored articles, because the interaction is fast and brand
|
|
feedback can be incorporated instantly. For content series and personal columns, I use Hermes, letting its writing
|
|
Skill grow over time. The two complement each other — no conflict.
|
|
|
|
|
|
===== 第 64 页 =====
|
|
§15 Multi-Agent Orchestration: Running Three Horses at
|
|
Once
|
|
Multi-Agent Orchestration: Running Three Horses at Once
|
|
When one agent isn't enough, put three to work in parallel. delegate_task is Hermes's most powerful tool —
|
|
and the easiest to misuse.
|
|
Why You Need Multiple Agents
|
|
A single agent's ceiling is determined by its context window and toolset. When a task is complex enough, a solo
|
|
agent hits two walls.
|
|
Context explosion. One agent handling research, coding, and testing at the same time — all that information
|
|
crammed into a single context, interfering with each other. Web page content from research eats up tokens, leaving
|
|
insufficient room for code reasoning.
|
|
Time bottleneck. Three tasks running sequentially at 30 minutes each means 90 minutes total. Run them in
|
|
parallel, and total time equals the slowest one.
|
|
Hermes's delegate_task tool exists to solve both problems. It can launch up to 3 sub-agents simultaneously, each
|
|
with its own independent context and toolset.
|
|
delegate_task in Detail
|
|
delegate_task isn't just "spawn a thread." It does several critical things.
|
|
Feature
|
|
Description
|
|
Independent
|
|
context
|
|
Each sub-agent has its own conversation history, preventing pollution of the main agent's context
|
|
Restricted toolset
|
|
You specify which toolsets a sub-agent can use. Additionally, delegate_task, clarify, memory,
|
|
send_message, and execute_code are always blocked for sub-agents
|
|
Isolated terminal
|
|
sessions
|
|
Each sub-agent has its own terminal, no interference
|
|
Max 3 concurrent
|
|
Hard-coded limit to prevent resource exhaustion
|
|
Result relay
|
|
When a sub-agent finishes, results are returned to the main agent for consolidation
|
|
The 3-concurrent limit is a deliberate design choice. Nous Research found in testing that beyond 3 sub-agents, the
|
|
main agent's consolidation quality drops sharply. It's not a compute limitation — it's an attention dispersion
|
|
|
|
|
|
===== 第 65 页 =====
|
|
problem when LLMs try to integrate too many independent information sources.
|
|
A Real Example: Competitive Analysis Report
|
|
Say you need to write a competitive analysis of AI coding tools, covering three products: Claude Code, Cursor, and
|
|
Hermes Agent. The traditional approach is researching them one by one, then manually integrating everything.
|
|
With delegate_task, the workflow becomes:
|
|
Main Agent
|
|
Split tasks + define template
|
|
→
|
|
Parallel execution
|
|
→
|
|
Main Agent
|
|
Consolidate + generate report
|
|
Zooming into the parallel execution:
|
|
1
|
|
Main agent defines the task template
|
|
"Research [product name] along these dimensions: positioning, core features, technical architecture, pricing,
|
|
community size, pros and cons. Output as a markdown table."
|
|
2
|
|
Dispatch three sub-agents
|
|
Sub-agent A researches Claude Code, sub-agent B researches Cursor, sub-agent C researches Hermes. Each
|
|
sub-agent gets only web and browser — two tools, no file or terminal needed.
|
|
3
|
|
Main agent consolidates
|
|
After all three sub-agents return their research, the main agent integrates everything into a comparison
|
|
report, adding its own judgment and recommendations.
|
|
Three research tasks in parallel — total time goes from "A+B+C" to "max(A, B, C)." In practice, a competitive
|
|
analysis that normally takes 40 minutes is done in 15.
|
|
|
|
|
|
===== 第 66 页 =====
|
|
Security Design of Restricted Toolsets
|
|
The restricted toolset in delegate_task isn't just an efficiency feature — it's a security mechanism.
|
|
Imagine this scenario: you send a sub-agent to search for a code snippet online, and it finds one with malicious
|
|
injection. If that sub-agent also has terminal permissions, it might execute the code. But if it only has web
|
|
permissions, the search result comes back as text for the main agent to review.
|
|
That's the value of restricted toolsets: the principle of least privilege, applied at the agent level.
|
|
推荐
|
|
Good practice: Research sub-agents get only
|
|
web+browser. Coding sub-agents get only
|
|
terminal+file+code_execution. Consolidation sub-agents
|
|
get no external tools at all — they only process text.
|
|
不推荐
|
|
Bad practice: Every sub-agent gets the full toolset.
|
|
Convenient, sure, but you lose the isolation that keeps
|
|
things safe.
|
|
Relationship to Anthropic's Three-Agent Architecture
|
|
Anthropic's agent design guide proposed a classic three-agent architecture: Planner for planning, Generator for
|
|
execution, Evaluator for assessment. This pattern has proven effective in many scenarios.
|
|
Hermes's delegate_task shares similarities with this architecture, but has key differences:
|
|
Dimension
|
|
Anthropic Three-Agent
|
|
Hermes delegate_task
|
|
Role assignment
|
|
Fixed roles (plan/execute/evaluate)
|
|
Task-driven, flexible roles
|
|
Communication
|
|
Chain (plan → execute → evaluate)
|
|
Star topology (main agent ↔ sub-agents)
|
|
Parallelism
|
|
Typically sequential
|
|
Up to 3 concurrent
|
|
Memory
|
|
No built-in memory
|
|
Main agent maintains full memory
|
|
Hermes's model is more flexible. You can use delegate_task to implement Anthropic's three-agent architecture
|
|
— one sub-agent for planning, one for execution, one for evaluation. But you can also have three sub-agents doing
|
|
the same type of parallel task (like researching three products simultaneously). The architecture isn't fixed — it
|
|
depends on how you decompose the task.
|
|
Anthropic's three-agent architecture is a mental framework, telling you "complex tasks can be split into planning,
|
|
execution, and evaluation." Hermes's delegate_task is an execution tool that turns that framework into reality.
|
|
One governs design; the other governs implementation.
|
|
|
|
|
|
===== 第 67 页 =====
|
|
核心建议
|
|
Multi-agent orchestration is easy to over-engineer. Not every task needs to be split into sub-agents. If a task fits
|
|
comfortably within a single agent's context window, splitting actually adds communication overhead and
|
|
consolidation errors. Only reach for delegate_task when context is insufficient or you need parallel speedup.
|
|
Rule of thumb: If you find yourself writing lengthy consolidation instructions for the main agent to integrate sub-
|
|
agent results, the task decomposition is probably wrong. Good decomposition makes consolidation simple: each
|
|
sub-agent's output should be self-contained, uniformly formatted, and directly composable.
|
|
|
|
|
|
===== 第 68 页 =====
|
|
§16 Hermes vs OpenClaw vs Claude Code: Not a Choice
|
|
Not a Choice — A Combination
|
|
These three tools aren't three roads. They're three horses. The question isn't which one to ride — it's figuring
|
|
out which one hauls cargo, which one covers distance, and which one guards the house.
|
|
Three Species
|
|
Over the past year, the AI Agent ecosystem has spawned too many tools to count. But the three genuinely worth a
|
|
serious look, in my opinion, are Claude Code, OpenClaw, and Hermes Agent.
|
|
Not because they're the most hyped, but because they represent three fundamentally different design
|
|
philosophies.
|
|
Claude Code is an interactive coding tool. You sit at the terminal, give it requirements, and it writes code, runs
|
|
tests, commits to git. You're present the whole time, like pair programming with a very capable engineer. The core
|
|
value is real-time code productivity.
|
|
OpenClaw is a "configuration-as-behavior" framework. You define the agent's personality, knowledge, and skills
|
|
entirely through SOUL.md and Skill files. The config files determine what the agent is. The core value is
|
|
predictability, auditability, and reproducibility.
|
|
Hermes Agent is an autonomous background engine. You deploy it on a server, and it runs 24/7 — remembering,
|
|
creating Skills, improving itself. The core value is autonomy and self-improvement.
|
|
Dimension
|
|
Claude Code
|
|
OpenClaw
|
|
Hermes Agent
|
|
Core
|
|
philosophy
|
|
Interactive coding
|
|
Configuration as behavior
|
|
Autonomous background
|
|
+ self-improvement
|
|
Your role
|
|
Sitting at the
|
|
terminal directing
|
|
Writing config files to define behavior
|
|
Deploy and check in
|
|
occasionally
|
|
Memory
|
|
mechanism
|
|
CLAUDE.md + auto-
|
|
memory
|
|
Multi-layer memory (SOUL.md + Daily Logs +
|
|
semantic search), transparent and
|
|
controllable
|
|
Three-layer self-improving
|
|
memory
|
|
Skill source
|
|
Manually installed
|
|
ClawHub 5,700+
|
|
Agent-created +
|
|
community Hub
|
|
Run mode
|
|
On-demand
|
|
On-demand
|
|
24/7 background
|
|
Deployment
|
|
Local CLI
|
|
(subscription)
|
|
Local CLI (free + API costs)
|
|
$5 VPS / Docker /
|
|
Serverless
|
|
|
|
|
|
===== 第 69 页 =====
|
|
See the distinction? These three tools aren't even solving the same problem.
|
|
|
|
|
|
===== 第 70 页 =====
|
|
Which Tool for Which Scenario
|
|
I've used Claude Code for half a year, OpenClaw for several months, and I've been tinkering with Hermes recently.
|
|
My takeaway: picking a tool isn't about which one is more powerful — it's about which interaction model fits your
|
|
scenario.
|
|
Scenario
|
|
Recommended Tool
|
|
Why
|
|
Building new features, refactoring
|
|
code
|
|
Claude Code
|
|
Needs real-time feedback and human judgment
|
|
Setting up standardized agents for a
|
|
team
|
|
OpenClaw
|
|
SOUL.md is transparent, auditable, reproducible
|
|
24/7 code review
|
|
Hermes
|
|
Cron scheduling + GitHub MCP, runs unattended
|
|
Personal knowledge assistant
|
|
Hermes
|
|
Three-layer memory accumulates across sessions, gets
|
|
smarter over time
|
|
Building a customer support /
|
|
community bot
|
|
Hermes
|
|
Native 12+ platform Gateway, multi-channel
|
|
Rapid product idea validation
|
|
Claude Code
|
|
Fast to start, fast to iterate, real-time course correction
|
|
Enterprise scenarios needing high
|
|
control
|
|
OpenClaw
|
|
Transparent config, predictable behavior
|
|
Long-term content creation project
|
|
Hermes + Claude
|
|
Code
|
|
Hermes for ongoing research and accumulation, Claude
|
|
Code for writing
|
|
That last row matters. Many scenarios can't be handled by a single tool. In a long-term content project, Hermes
|
|
handles daily automated information gathering and memory accumulation, while Claude Code sits down and
|
|
actually writes the piece. Each owns a different leg of the relay.
|
|
Convergence or Divergence
|
|
Here's an interesting pattern: these three tools are learning from each other's strengths.
|
|
Claude Code added auto-memory, moving toward Hermes-style persistent memory. OpenClaw's ClawHub has
|
|
5,700+ community Skills; Hermes is building its own Skill Hub. Hermes supports the agentskills.io standard,
|
|
meaning it can directly use Skills from the Claude Code ecosystem.
|
|
Looks like convergence. But I think the underlying divergence is actually widening.
|
|
Claude Code is fundamentally about real-time conversation between human and AI. No matter how much
|
|
memory and automation it adds, the fact that you're sitting there watching it work won't change. Anthropic's
|
|
business model dictates this: subscription-based, charged by your usage time.
|
|
|
|
|
|
===== 第 71 页 =====
|
|
Hermes is fundamentally about AI running autonomously in the background. No matter how many interactive
|
|
interfaces it adds, its core value is that it keeps working when you're not there. The MIT open-source + self-hosted
|
|
model dictates this.
|
|
OpenClaw sits in the middle. It doesn't emphasize real-time interaction like Claude Code, nor autonomous
|
|
operation like Hermes. Its unique value is "transparent and controllable." SOUL.md lets you see at a glance exactly
|
|
what an agent will and won't do. For enterprise compliance scenarios, this property is irreplaceable.
|
|
|
|
|
|
===== 第 72 页 =====
|
|
agentskills.io: Why Skill Interoperability Matters
|
|
In early 2026, the agentskills.io standard started gaining adoption across multiple tools. Currently 16+ tools
|
|
support it, including Claude Code, Cursor, OpenAI Codex, Gemini CLI, and Hermes.
|
|
What does this mean?
|
|
It means a Skill you wrote for Claude Code can be directly used by Hermes. A Skill auto-created inside Hermes can
|
|
be brought into Claude Code. Skills are no longer tied to a specific tool — they become portable capability units.
|
|
The long-term impact of this might be bigger than any individual tool. Because it's saying: no matter which horse
|
|
you ride, the saddle is universal.
|
|
The time you invest in writing Skills won't be wasted if you switch tools. Your Skill library is your own asset, not a
|
|
platform's appendage.
|
|
OpenClaw's ClawHub has 5,700+ Skills. If those Skills can be called directly by Hermes through the agentskills.io
|
|
standard, Hermes's capability boundary expands instantly. Conversely, Skills that Hermes auto-creates and
|
|
improves can feed back into the broader ecosystem.
|
|
Not a Choice — A Combination
|
|
Having written three orange books (Claude Code, Harness Engineering, OpenClaw), my biggest takeaway about
|
|
this space is: the winner won't be any single tool — it'll be the people who know how to combine tools.
|
|
My own workflow is already a combination. Claude Code handles everything that needs me present: writing
|
|
articles, writing code, making product decisions. It's my "day shift."
|
|
Hermes (or similar autonomous agents) handles what doesn't need me present: monitoring repos, running
|
|
scheduled research, maintaining knowledge bases. It's my "night shift."
|
|
OpenClaw's SOUL.md and Skill system give me a standardized configuration language. Whether Claude Code or
|
|
Hermes is running underneath, the behavioral constraints are written the same way.
|
|
HuaShu's take: Don't "choose" between these three tools. Ask yourself three questions: Which tasks need me
|
|
watching? Which tasks can run in the background? Which scenarios need transparent auditability? The answers
|
|
naturally sort the tools into their respective positions.
|
|
Competition among agent tools won't converge to a single winner. Just like you don't use a hammer to turn a
|
|
screw — interactive coding, configuration management, and autonomous operation are three distinct work modes
|
|
that will coexist long-term.
|
|
The truly interesting question isn't "which is better" but "how do we make them collaborate." agentskills.io is
|
|
already paving that road.
|
|
|
|
|
|
===== 第 73 页 =====
|
|
§17 The Boundaries of Self-Improving Agents: How Far
|
|
Can They Go
|
|
The Boundaries of Self-Improving Agents
|
|
Hermes's most exciting capability is also its most unsettling one.
|
|
Revisiting That Diagram
|
|
In the final chapter of the Harness Engineering orange book, I wrote about the diagram Kief Morris drew. Three
|
|
layers: in the loop, on the loop, out of the loop.
|
|
In the loop: reviewing every line of the agent's output. On the loop: not checking outputs, just holding the reins.
|
|
Out of the loop: you say what you want, the agent handles everything.
|
|
My conclusion then was that on the loop is probably the best balance. You're not duplicating effort, but you're still
|
|
there.
|
|
Hermes Agent pushes this discussion to a new place.
|
|
Its learning cycle is automatic. It creates Skills on its own, improves Skills on its own, decides what to remember on
|
|
its own. After you deploy it, it keeps getting stronger. This isn't on the loop anymore. You don't even have to adjust
|
|
the reins — the reins are growing by themselves.
|
|
Is this progress or risk?
|
|
Can Skill Self-Improvement Go Out of Control
|
|
Let's start with the technical layer. Hermes's Skill self-improvement has several constraints.
|
|
Skill files are readable markdown. Not black-box neural network weights — they're text you can open and read.
|
|
What it changed, you can see in the diff.
|
|
Memory data is local. Built on SQLite + FTS5, data sits on your local disk. You can inspect and delete it directly.
|
|
There's no "the agent secretly learned something you don't know about" situation.
|
|
Tool permissions are sandboxed. The agent can't arbitrarily acquire new system permissions; the toolset must be
|
|
explicitly configured.
|
|
Technically speaking, Hermes's self-improvement is controlled and auditable. You can see what it changed, roll it
|
|
back, delete it.
|
|
But technically controlled doesn't mean practically controlled.
|
|
|
|
|
|
===== 第 74 页 =====
|
|
The problem is on the human side.
|
|
Are you really going to check which Skills the agent modified every day? Are you going to audit its memory
|
|
database? Probably not. The whole appeal of deploying Hermes is "not having to babysit it." If you had to review
|
|
its self-improvement results daily, how would that be different from maintaining Skills manually?
|
|
This contradiction is fundamental: the value of an autonomous agent lies in not having to watch it, but safety
|
|
requires you to watch it.
|
|
Kief Morris's insight holds true again here: the difference between in the loop and on the loop becomes most
|
|
obvious when you're unhappy with the result. But if the agent self-improved a Skill and you never noticed
|
|
anything wrong, when would you ever catch it?
|
|
Nous Research's Choice
|
|
The founding team at Nous Research made a clear choice on this issue: user control first.
|
|
They've explicitly stated that models should be "steerable" — users can adjust behavior as needed, unconstrained
|
|
by corporate content policies.
|
|
This isn't just talk. Hermes Agent's MIT license means you own the entire source code. You can audit every step of
|
|
the learning cycle's logic, modify the thresholds, frequency, and scope of self-improvement, or turn off automatic
|
|
Skill creation entirely.
|
|
Compared to closed-source agents, this transparency gives you a floor: in the worst case, you can see what all the
|
|
code is doing.
|
|
But let me say something a bit uncomfortable.
|
|
"You can see the code" and "you've actually read the code" are two very different things. The vast majority of users
|
|
won't read source code. The vast majority of people deploying Hermes will use default settings. The MIT license
|
|
gives you the right to audit, but it doesn't guarantee you'll exercise that right.
|
|
|
|
|
|
===== 第 75 页 =====
|
|
Open Source vs Closed Source: Different Shapes of Trust
|
|
This leads to a bigger question: when it comes to self-improving agents, is open source or closed source more
|
|
trustworthy?
|
|
The intuitive answer is open source. Transparent code, community audits, MIT license.
|
|
But reality is more nuanced than intuition.
|
|
Claude Code is closed source. You don't know what Anthropic's agent internals look like. But Anthropic has a clear
|
|
commercial incentive to keep agent behavior predictable: if an agent goes rogue and damages a user's codebase,
|
|
they lose subscribers. Business pressure is a form of constraint.
|
|
Hermes is open source. You can see all the code. But if the agent's self-improvement causes a problem, Nous
|
|
Research has no commercial obligation to fix it for you. The other side of the MIT license: you bear the
|
|
consequences.
|
|
Two shapes of trust: one is "I trust your business incentives," the other is "I trust my own ability to audit."
|
|
For people with technical chops, open source is clearly better — you control everything. For people who don't
|
|
want to touch code and just want to use the tool, a closed-source commercial service might actually be safer,
|
|
because someone else is on the hook.
|
|
Where Is the Ceiling of Self-Improvement
|
|
Back to the core question: how far can self-improving agents go?
|
|
My assessment: the ceiling isn't technical — it's the feedback signal.
|
|
Hermes's self-improvement loop relies on a key assumption: it can judge whether its own improvements are good
|
|
or bad. It modifies a Skill, the next task goes better — that's positive feedback. But who defines "better"?
|
|
If you're there giving feedback, the loop works. You say "that's wrong," it adjusts. This is supervised improvement.
|
|
If you're not there, the agent can only use its own evaluation criteria. It thinks the response was faster, more
|
|
accurate. But "fast" and "accurate" don't equal "correct." Some errors require domain knowledge to catch. The
|
|
agent doesn't know what it doesn't know.
|
|
In the Harness Engineering book, I wrote: Mitchell Hashimoto could write excellent harness for Ghostty because he
|
|
understood every detail of terminal emulators. A self-improving agent doesn't have Mitchell's domain knowledge.
|
|
It can optimize execution efficiency, but it can't judge whether the direction is right.
|
|
Self-improvement makes agents run faster in a known direction. But the direction itself still needs a human to
|
|
set.
|
|
|
|
|
|
===== 第 76 页 =====
|
|
Questions for You
|
|
At this point, I'm not going to wrap things up with a neat conclusion. Because these questions don't have definitive
|
|
answers, and the answers will keep shifting as technology evolves.
|
|
But I want to lay out a few questions worth thinking about over time.
|
|
How much autonomous self-improvement are you comfortable with?
|
|
Rewriting Skill files? Sure. Auto-creating new Skills? Probably fine. Modifying its own core configuration? Maybe
|
|
not. Modifying its own learning loop logic? Absolutely not. Where do you draw the line?
|
|
Who audits the results of self-improvement?
|
|
You do it yourself — how often? The community audits — do you trust the community's judgment? Nobody audits
|
|
— are you okay with that risk?
|
|
Do self-improving agents need a "forgetting" mechanism?
|
|
Humans forget, and that's not a bug — it's a feature. Outdated experience fading away is what keeps it from
|
|
polluting current judgment. Agent memory only grows; patterns learned three months ago might be obsolete. Who
|
|
tells the agent "this one should be forgotten"?
|
|
What does Kief Morris's concern look like in the Hermes context?
|
|
Morris worried: if junior developers never touch code details, who will design the harness in the future? In the
|
|
Hermes context, the question becomes: if the agent designs its own reins, who judges whether the reins are
|
|
designed correctly?
|
|
Maybe the answer is: you always need a human in the loop at some level. Not reviewing every line of code, but
|
|
understanding what the system is doing and why it's doing it.
|
|
Maybe the answer is: we don't know yet.
|
|
HuaShu's take: Self-improving agents are the most exciting direction in this space, but their ceiling is determined
|
|
by the degree of human involvement. A fully hands-off self-improving agent will win on efficiency but lose on
|
|
direction. The sweet spot might be: let the agent self-improve on the "how," while you own the "what" and the
|
|
"don't." That's not being lazy — it's a different kind of on the loop.
|
|
|
|
|
|
===== 第 77 页 =====
|
|
Hermes Agent: The Complete Guide
|
|
Orange Book Series
|
|
HuaShu · AI进化论-花生
|
|
A practical guide to the Nous Research open-source AI Agent
|
|
framework
|
|
From installation to multi-Agent orchestration
|
|
All Orange Books → huasheng.ai
|
|
Bilibili
|
|
·
|
|
WeChat: 花叔
|
|
·
|
|
X/Twitter
|
|
·
|
|
YouTube
|
|
·
|
|
Xiaohongshu
|
|
·
|
|
Website
|
|
Created by HuaShu · v260408 · April 2026
|
|
For learning purposes only. AI tools evolve rapidly — refer to official docs.
|