SkillOpt
A text-space optimizer that trains reusable natural-language skills for a frozen language model agent, refining a skill file through trajectory-driven edits and validation gates instead of updating model weights.
SkillOpt trains agent skills the way a neural network is trained, with epochs, batch size, learning rates and validation gates, but by editing natural-language skill text instead of model weights. Its training loop runs rollout, reflection, aggregation, selection, update and evaluation steps, and produces a deployable best_skill.md file.
SkillOpt-Sleep is a nightly, offline companion that reviews past agent sessions, replays recurring tasks, and consolidates validated skill updates behind a held-out validation gate, so an agent's skills improve between sessions without manual editing. It supports multiple model backends, including OpenAI, Azure, Claude, Qwen and MiniMax, ships with built-in benchmarks and a web dashboard, and integrates with Claude Code, Codex, Copilot, Devin and OpenClaw.
It installs from PyPI and is written in Python, under the MIT license.