An evolutionary self-improvement system for Hermes Agent that automatically optimizes the agent's skills, tool descriptions, and system prompts using DSPy and GEPA (Genetic-Pareto Prompt Evolution). Rather than GPU training, it works entirely through API calls: it reads the current skill or prompt, generates an evaluation dataset (synthetic or drawn from real session history from Claude Code, Copilot, and Hermes), mutates the text into candidate variants, evaluates them against execution traces, and selects the best. GEPA's distinguishing trait is that it reads traces to understand why a variant fails, then proposes targeted improvements. The README cites roughly $2-10 per optimization run.
The work is phased: skill-file (SKILL.md) evolution is implemented, while tool descriptions, system-prompt sections, tool implementation code (via the external Darwinian Evolver), and a continuous improvement loop are planned. Guardrails constrain every evolved variant: the full test suite must pass, size limits apply (skills at most 15KB, tool descriptions at most 500 characters), variants must stay cache-compatible and semantically faithful to the original purpose, and all changes land as pull requests for human review rather than direct commits.
