Agentic AI Atlas

II.

Page reference

page:docs-hermes-research-raw-self-evolution

Reading · 2 min

Hermes Agent Self-Evolution reference

This repository contains an evolutionary self-improvement system for Hermes Agent. The project uses DSPy combined with GEPA (Genetic-Pareto Prompt Evolution) to automatically refine and optimize agent skills, tool descriptions, system prompts, and code through reflective evolutionary search.

Pagewiki/docs/hermes-research/raw/self-evolution.mdOutgoing · 0Incoming · 1

Hermes Agent Self-Evolution

Source: https://github.com/NousResearch/hermes-agent-self-evolution

Overview

A notable feature is that "no GPU training required" -- the system operates entirely through API calls, with optimization runs costing approximately $2-10 each.

How It Works

The system follows a pipeline:

1. Read current skill/prompt/tool definitions 2. Generate evaluation datasets from execution traces 3. Run GEPA optimizer to create candidate variants 4. Evaluate candidates against constraint gates (tests, size limits, benchmarks) 5. Select the best variant for submission as a pull request

GEPA distinguishes itself by analyzing *why* failures occur, not merely registering that they happened, enabling targeted improvements.

What Gets Optimized

Phase	Target	Engine	Status
1	Skill files (SKILL.md)	DSPy + GEPA	Implemented
2	Tool descriptions	DSPy + GEPA	Planned
3	System prompt sections	DSPy + GEPA	Planned
4	Tool code	Darwinian Evolver	Planned
5	Continuous improvement	Automated pipeline	Planned

Key Guardrails

Every evolved variant must pass:

Full test suite (100% pass rate)
Size constraints (skills <=15KB, descriptions <=500 chars)
Caching compatibility checks
Semantic preservation validation
Human code review before deployment

Installation & Usage

bash

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

python -m evolution.skills.evolve_skill \
    --skill github-code-review \
    --iterations 10 \
    --eval-source synthetic