Agent evolution

A way to make an AI agent measurably better at something: keep a population of complete, standalone agents, breed new ones with operators that start from a diagnosis rather than a guess, freeze every agent the moment it has a score, and judge them all against ground truth with the controls that stop you fooling yourself.

• docs/practice.md, the governing doc.
• implementation/, the implementation of that description: population.md (what an agent is on disk), evolution.md (the operators, selection, where proposals come from), metric.md (what a project must supply and what makes a score worth optimising), researching.md (the rules a research agent works under), operating.md (dispatch, the record, asks, the rendering).
• skill/, an installable skill for AI agents: copy or symlink it into an agent's skills directory.
• tools/, project-agnostic: check-tree validates a population, new-node scaffolds an agent, render-tree draws the page.
• index.html, the human-readable representation of ./docs.

Using it in a project

Drop a .agent-evolution.json at the project root naming where its population, scores and asks live, then run the tools against it:

{
  "name": "trader-agent",
  "population": "traders",
  "scores": "state/bench",
  "asks": "asks",
  "metric": {"name": "points from the outcome", "lowerIsBetter": true}
}

License

MIT

---

This repo uses ddd-practice: docs are the source of truth, everything else is derived. ./docs holds one governing document, practice.md, elaborated by ./implementation. The human-readable representation is index.html at the repo root.