Your coding agent's failures are now a curriculum

What it does

EvoSkill runs a mutation-and-evaluation loop on coding agents. It treats failed trajectories as training material, proposing new skills and prompt variants, testing them on held-out tasks, and promoting the best combinations into an entirely new agent program. The result is an evolved configuration—system prompts, tools, and reusable skill folders—that drops into Claude Code, Codex CLI, OpenHands, Goose, and others.

The interesting bit

Where its predecessor GEPA revised a single prompt in place, EvoSkill evolves the whole agent program. It jointly mutates skills and prompts, then packages the survivors as portable folders with metadata and helper scripts that can hop across both different agent harnesses and different LLMs.

Key highlights

• Compatible with Claude Code, Codex CLI, OpenCode, OpenHands, Goose, and Harbor out of the box.
• Skills evolved on one model can reportedly boost performance on others (cross-model transfer), and a skill learned for one task can help on unrelated tasks (cross-task transfer).
• Integrates with Harbor for containerized benchmarks with built-in verifiers, or works with plain CSV question/answer datasets.
• Each iteration produces a new agent program on a dedicated git branch, making the evolution auditable and rollback-friendly.
• Provider-agnostic: works with OpenRouter, Anthropic, OpenAI, Fireworks, and supports models from Claude to Qwen.

Caveats

• Two major features—evolution without a benchmark and continuous evolution—are explicitly flagged as unfinished (🛠️).
• OpenHands integration relies on a fallback JSON extraction workaround due to missing native structured output.
• Agent support is uneven: OpenCode and Goose require recent CLI versions, and Codex CLI needs a symlink-based skill discovery hook.

Verdict

Worth a look if you’re running coding agents at scale and want automated, data-driven prompt engineering without vendor lock-in. Probably overkill if you’re just occasionally asking Claude to refactor a function.