🧬 cap-evolve

cap-evolve improves an AI agent’s prompts, tools and skills by learning from failed evaluation traces. You bring the agent and the eval you already have; cap-evolve runs the loop — evaluate → diagnose failures → propose an edit → keep it only if it beats a held-out split → commit — and reports one honest number. It optimizes what your agent reads, not its weights.

Results at a glance

Benchmark	Split	Baseline → Optimized
τ²-bench airline (policy + tools)	sealed test	30.0 → 47.5 (+58%)
SkillsBench (skill package)	sealed test	0.556 → 0.667 (+20%)
RH-SWE-bench (skill + system prompt)	val, fit metric	0.580 → 0.765 (+32%)

See the project’s RESULTS.md for methodology and caveats.

Highlights

• More than prompts — system prompts, executable tool code, MCP tool surfaces and whole skill packages, alone or jointly.
• Learns from real failures — reads full trajectories and per-task causal feedback, so edits don’t regress past wins.
• Honest evaluation — a val-only significance gate accepts edits; the test split is sealed and scored once.
• Inspectable — every candidate is a git commit; the dashboard shows cost, diffs, lineage and a pass/fail heatmap.