wikis / Hermes / wiki / entities / hermes-self-evolution.md view as markdown
Overview
hermes-agent-self-evolution (github.com/NousResearch/hermes-agent-self-evolution, 893 stars per get-hermes.ai, 2026-04-12) is an official nous research project: an evolutionary self-improvement pipeline using DSPy and GEPA (Genetic Evolution of Prompt Architectures). The awesome-hermes-agent list describes it as "the research pipeline for optimizing Hermes's own prompts and behaviors."
It is the offline/research-grade counterpart to the self improvement loop that runs inside every Hermes session. The in-agent loop (the "GAPA" mechanism community videos describe as "back propagation but for prompts instead of model weights") fires continuously during normal use; this project packages the same idea as a standalone evolutionary optimization pipeline you run deliberately — e.g., on a schedule against your agent's accumulated trajectories.
Naming caution: the 24/7 Self-Evolving video spells the in-agent mechanism "GAPA," while this project's documentation says "GEPA." transcript 247 self evolving treats them as the same approach. The exact relationship between the in-session loop and this repo's pipeline is inferred, not confirmed from the repo itself — low confidence on that mapping.
Characteristics
- Repo: https://github.com/NousResearch/hermes-agent-self-evolution
- Maintainer: Nous Research (official project, not community)
- Stars: 893 (get-hermes.ai community page, 2026-04-12)
- Approach: DSPy + GEPA — evolutionary optimization of prompts/behaviors rather than weight updates
- Scope: optimizes the Hermes harness's own prompts and behaviors; complements (does not replace)
tinker-atropos, which does actual RL fine-tuning of model weights on agent trajectories (see ml research pipeline) - Listed in: community awesome hermes under Official Resources, and on the get-hermes.ai community page under Official Nous Research Projects
How to Use
Concrete setup commands are not present in the sources ingested so far — the steps below are the operational pattern recommended by the awesome-hermes-agent "Operational Playbooks" section, not a verified walkthrough:
- Clone https://github.com/NousResearch/hermes-agent-self-evolution and follow the repo README.
- Nightly self-evolution + guardrail evaluation — run the pipeline on a schedule (Hermes cron), then run a second verification cron to score output quality and block optimization-loop gaming. This guardrail step is explicitly recommended; unguarded self-optimization can reward-hack its own metric.
Until the repo README is ingested as a raw source, treat any specific CLI invocation as unverified.
Related Entities
- nous research — maintainer; this is one of four official ecosystem projects alongside hermes paperclip adapter and autonovel
- self improvement loop — the in-session GAPA loop this project industrializes
- ml research pipeline — trajectory capture and RL training infrastructure that feeds and complements it
- skills system — skills are the substrate that prompt-evolution outputs land in
- transcript 247 self evolving — community framing of the GAPA/GEPA mechanism
- community awesome hermes — source of the nightly-run operational playbook