evolving-ai-agents

What problem does it solve?

This Skill helps you automatically evolve AI agents so their prompts, skills, and memory improve over time against measurable benchmarks instead of relying on slow, manual trial-and-error.

Core Features & Use Cases

• Benchmark-driven evolution loop: Runs solve → observe → evolve → gate → reload cycles using benchmark feedback to guide targeted mutations.
• File-system-based agent optimization: Represents evolving agent state (prompts, skills, memory, tools) as versioned workspace files so changes are explicit, inspectable, and rollbackable.
• Automated skill discovery and refinement: Learns reusable procedures by proposing, curating, merging, and updating skills based on observed failures and performance changes.
• Git snapshot gating for safety: Uses git snapshots for rollback-friendly gating so rejected or harmful mutations don’t linger.

Quick Start

Use A-Evolve to run 10 evolution cycles for a SWE-bench agent by giving the command: pip install a-evolve && python -c "import agent_evolve as ae; evolver=ae.Evolver(agent='swe', benchmark='swe-verified'); r=evolver.run(cycles=10); print(r.final_score)"