Hermes Agent: An AI Agent That Writes Its Own Skills

An ordinary agent solves a task and forgets how it did it. Next time the task comes up, it thinks from scratch again, spending the same tokens and the same time. The open agent Hermes Agent from Nous Research works differently: after a complex task it writes down a skill for how to do it, and next time it uses the ready solution.

How the learning works inside

The developers describe it as a built-in learning loop. The agent creates skills from its own experience and improves them on reuse. It reminds itself to save important knowledge, searches its own past conversations, and gradually builds a model of who you are and what you usually need, across different sessions.

The skills Hermes Agent writes for itself are compatible with the open agentskills.io standard. This means the solutions the agent accumulates aren't locked inside one installation. They can be moved or used in other tools that support the same standard.

What is actually inside

The project is released under the MIT license, so the open code can be used and changed freely. At the time of checking, the repository has more than 250 thousand stars on GitHub and over 54 thousand forks. It is one of the fastest-growing open agent projects of 2026.

Inside there is a full terminal interface with command autocomplete. You can connect through Telegram, Discord, Slack, WhatsApp, Signal, or the command line, and everything comes into a single gateway. A built-in scheduler runs tasks on a schedule (cron) for unattended background work.

The agent is compatible with Nous Portal, OpenRouter, OpenAI, and your own model endpoints. It can run on a 5-dollar-a-month VPS as well as on GPU clusters or serverless platforms like Modal and Daytona.

Why this matters to someone who already builds their own agents

If you already have working AI agents for content or business tasks, the idea of self-written skills solves a specific problem: usually the knowledge of how the agent solved a task last time lives either in the head of the person who set it up, or nowhere at all. Hermes Agent makes this step automatic: a skill is written and stored without manual intervention.

This is not a reason to rewrite an already working agent system from scratch. But if a task repeats regularly and gets solved a little differently each time, it's worth looking at whether you can build a similar mechanism of self-recording the solution into at least one specific repeating process, instead of keeping all the knowledge in one person's memory.

How to install

Installation takes one command in the terminal:

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

After installation the agent starts with the hermes command, and then model setup and messenger connection go through a dialog inside the terminal itself.

I write about working agent setups and content automation in the channel «Растём с ИИ» ("Growing with AI"). A related piece: how to write a skill for Claude Code so it really works. If you need to go through your own system of agents and skills in person, you can start with a free consultation.