AI Agent framework with more than 210,000 Stars: Why can Hermes Agent evolve itself?
Star AI Agent framework with over 210,000 users: Why can Hermes Agent evolve itself?
I recently spent a week digging into Hermes Agent, an open source project from NousResearch. It has surpassed 212,000 GitHub stars, over 39,000 forks, and over 370 community contributors. v0.18.0 alone incorporates nearly 1,000 PRs and contains approximately 251,000 lines of new code.
This number is real in the open source AI community.
But what interests me more is not the numbers, but why so many people are willing to invest time in this project.
What exactly is Hermes Agent?
Simply put, Hermes Agent is a self-evolving AI Agent framework. Its core positioning is: "The agent that grows with you".
This may sound like a marketing slogan, but if you look closely at its architecture, you will find that "self-evolution" is not a metaphor, but the actual behavior of the system in operation.
It’s not another LLM wrapper
Many AI Agent projects are essentially a "wrapper" - you enter a prompt, it calls the LLM API and returns the results. There is nothing wrong with this pattern per se, but it also means:
- You need to manually tell it "how to do the last task" every time
- It does not learn from its mistakes
- It doesn't get better with use
Hermes Agent does things differently. It has a Closed Loop Learning System:
1. Automatically create Skill: When a complex task is completed, it will automatically extract experience and create a reusable Skill.
1. Skill Self-Improvement: In subsequent use, it will automatically improve the Skill based on actual performance.
1. Experience accumulation: These skills will be stored persistently and become cross-session knowledge.
This means that the more you use it, the better it understands you.
Core architecture analysis
1. Self-evolving closed-loop system
The learning mechanism of Hermes Agent is its most unique feature. Let me illustrate with a concrete example:
Suppose you ask it to automate a file processing process for you. The first time, it may require you to tell it each step step by step. But when the task is completed, it will:
- Analyze the process of this mission
- Extract key decision points and best practices
- Create a reusable Skill
- Automatically call this Skill the next time you encounter a similar task
The key is: if the Skill is found not to be good enough on the next mission, it will automatically improve it. This process is automatic and continuous.
2. Mixture-of-Agents (MoA)
One of the biggest new features in v0.18.0 is Mixture-of-Agents becoming first-class citizens.
What is a MoA? To put it simply, it allows you to call multiple LLMs at the same time, let them give their own answers, and then an "aggregation model" combines all the answers to give the final answer.
This may sound a bit complicated, but it's very intuitive to use:
- You can choose the MoA preset "my-council"
- It will automatically call multiple models (such as Claude, GPT, Gemini)
- Each model's answer is shown to you
-Finally give a comprehensive answer
Why is this important?
Because some questions really require multiple perspectives to give the best answer. For example:
- A technical architecture issue may require an architect’s perspective + an engineer’s perspective + a security expert’s perspective
- A creative writing assignment may require the poet’s perspective + the editor’s perspective + the target reader’s perspective
MoA makes this possible without requiring manual coordination on your part.
3. Multi-platform life
Hermes Agent doesn't just run in the terminal. It can:
- Telegram: Talk to it in Telegram
- Discord: used in Discord server
- Slack: integrated into your Slack workspace
- WhatsApp: Interact with it via WhatsApp
- CLI: traditional command line interface
All of this is managed through a Gateway process. This means you can start a task in Telegram and then switch to Discord to continue it without losing context.
4. Scheduling and automation
Hermes Agent has a built-in scheduling system (cron) that can:
- Perform tasks on a regular basis (e.g. report every morning)
- Automatic backup
- Regular audits
These tasks can be described in natural language without the need to write programming code. For example:
# Every morning at 9 am, check my GitHub notifications and notify me if there are new PRs
hermes cron schedule "Check GitHub notifications every morning at 9 am" --platform telegramIt will automatically create a cron job and execute it at the specified time.
5. Subagents and parallelization
For complex tasks, Hermes Agent can create sub-agents for parallel processing:
- A subagent is responsible for searching information
- Another subagent is responsible for analyzing the data
- A third subagent is responsible for writing the report
All results are then compiled. This means you can work on multiple tasks simultaneously without the need for manual coordination.
Why is it worthy of attention?
1. It is open source, and truly open source
Although many AI Agent projects are open source, their core functions require paid APIs. Hermes Agent is different:
- All features are open source
- You can use any LLM provider (OpenAI, Anthropic, OpenRouter, even local models)
- No vendor lock-in
2. Its learning system is real
As mentioned before, its self-evolution is not marketing talk, but an actual working system. This means:
- The longer you use it, the better it understands you
- Mistakes are automatically converted into improvements
- Experience will be persisted and will not be lost due to restarting
3. Its community is very active
370+ contributors, nearly 1,000 PRs per week, and 212k+ stars, these numbers show that this is not a "one-person project." It has a real, active community maintaining and improving it.
4. It is not designed for specific scenarios
Many AI Agent frameworks focus on specific scenarios (e.g. customer service, data analysis, code generation). Hermes Agent is different:
- It can handle any task
- It can learn how to handle new types of tasks
- It is not locked to a specific area
How to get started
If you are interested in Hermes Agent, here are the steps to get started:
1. Installation
Depending on your operating system, choose the appropriate installation method:
Linux/macOS/WSL2/Termux:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bashWindows(PowerShell):
iex (irm https://hermes-agent.nousresearch.com/install.ps1)The installer automatically handles all dependencies: uv, Python 3.11, Node.js, ripgrep, ffmpeg, and a Portable Git Bash (based on MinGit, about 45MB, will not conflict with system Git).
2. Initialization
source ~/.bashrc # Reload shell
hermes # Get started!3. Start creating Skill
After you complete your first complex task, try:
/hermes learnThis triggers the learning process, which analyzes the task and builds a Skill.
4. Try MoA
During model selection, select a MoA preset (e.g. "my-council") and try it out on complex problems.
My opinion
Hermes Agent is one of the products closest to the concept of "AI assistant" that I have ever seen.
It's not another "enter prompt, output results" tool. It is a system that will grow, learn, and adapt.
But I also have some doubts:
1. Does self-evolution introduce risks? What are the consequences if a Skill is improved incorrectly?
1. Does closed-loop learning require a large amount of historical data? For new users, the learning curve may be longer.
1. How safe is it? When an agent can act autonomously, how to ensure that it does not perform harmful actions?
There are no standard answers to these questions, but it's worth discussing.
Conclusion
Hermes Agent represents an important direction of AI Agent: not a passive tool, but an active partner.
It doesn't end just because you use it once. It remembers your preferences, ways to improve it, and gets better over time.
If you're looking for an AI Agent that will grow, Hermes Agent is worth your time.
References