Hermes Agent Self-Evolution for Hermes Agent

Optimizes Hermes Agent skills with DSPy and GEPA through API calls, with no GPU training required

What Hermes Agent Self-Evolution does

Hermes Agent Self-Evolution reads an existing skill, prompt or tool description, generates an evaluation dataset, and runs the GEPA optimizer (Genetic-Pareto Prompt Evolution) over it. GEPA reads execution traces to see why something failed, then proposes targeted changes. Candidate variants are evaluated, and the best one is proposed as a pull request against the hermes-agent repository. Evaluation data can be synthetic or taken from real session history from Claude Code, Copilot and Hermes.

Everything runs through API calls, and the README estimates roughly $2 to $10 per optimization run. Each variant must pass the full pytest suite, size limits (skills up to 15KB, tool descriptions up to 500 characters), caching compatibility and a semantic-preservation check, and every change goes through human review rather than a direct commit. Only Phase 1, optimizing SKILL.md files, is implemented; tool descriptions, system prompts, tool code and a continuous loop are planned.

Key features

Evolves SKILL.md files with DSPy and GEPA (Phase 1, implemented)
Synthetic evaluation data or real session history through --eval-source
Reads execution traces to propose targeted mutations
Constraint gates: tests, size limits, caching compatibility, semantic preservation
Best variant is proposed as a PR for human review
Planned phases for tool descriptions, system prompts and tool code

When to use it

Improve an existing skill such as github-code-review using synthetic evaluation data
Tune a skill against your own Claude Code, Copilot and Hermes session history
Produce a reviewable pull request containing an optimized version of a Hermes Agent skill

Who it is for: Hermes Agent developers and skill authors who want to optimize skills with an automated, test-gated process.

How it fits with Hermes Agent

Published by Nous Research for Hermes Agent; it points at a local hermes-agent repository and targets its skills, prompts and code.

How to install Hermes Agent Self-Evolution

These commands are copied from the project's README. Check the repository for the latest steps before you run them.

git clone https://github.com/NousResearch/hermes-agent-self-evolution.git
cd hermes-agent-self-evolution
pip install -e ".[dev]"
export HERMES_AGENT_REPO=~/.hermes/hermes-agent

Requirements: A local hermes-agent repository set in HERMES_AGENT_REPO and API access for model calls; no GPU needed

Note: Only Phase 1, skill file optimization, is marked as implemented; the other four phases are listed as planned.

FAQ

What is Hermes Agent Self-Evolution?

Hermes Agent Self-Evolution is a Nous Research project that uses DSPy and GEPA to evolve and optimize Hermes Agent skills through reflective evolutionary search, without GPU training.

How do I install Hermes Agent Self-Evolution?

Clone the repository, run pip install -e ".[dev]" inside it, and set HERMES_AGENT_REPO to your hermes-agent checkout, for example ~/.hermes/hermes-agent. Then run python -m evolution.skills.evolve_skill with a skill name, an iteration count and an eval source.

How much does an optimization run cost?

The README estimates about $2 to $10 per optimization run. No GPU is needed, because everything runs through API calls.