evolving-ai-agents

To improve AI agent prompts automatically, run automated solve-observe-evolve cycles that mutate versioned workspace files based on scored benchmark task trajectories, refining prompts and memory instead of manual iteration.

Agent evolution uses LLM-driven workspace mutation to represent agent state as versioned files, applying mutations guided by feedback to refine prompts and skills over repeated cycles.

You need an agent implementing BaseAgent.solve and reload_from_fs, a BenchmarkAdapter with get_tasks and evaluate, and an evolution engine that applies mutations with gating and git snapshots.

Yes, the evolution engine discovers reusable procedures during the solve-observe-evolve cycles and curates or merges them to reduce skill redundancy in the versioned workspace.

Run the evolution engine for 10 cycles against your selected benchmark to learn from successes and failures, then save the optimized workspace containing the refined prompts and memory.

The approach applies to both single- and multi-tool agents, but requires strict gating and git snapshots to ensure mutations from benchmark feedback reliably improve task trajectories.