SkillOpt is a Python-based framework that trains reusable agent skills as editable text documents (best_skill.md) using a validation-gated optimization loop, with a CLI and optional WebUI. Latest releases include SkillOpt-Sleep, an offline self-evolution engine.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
SkillOpt is a text-space optimizer that trains reusable natural-language skills for frozen LLM agents through trajectory-driven edits, validation-gated updates, and deployable best_skill.md artifacts.
How it works
The tool treats the skill document as the trainable state of a frozen agent and uses an optimizer model to turn rollout scores into bounded edits (add/delete/replace) to a single skill document. A textual learning-rate budget, a rejected-edit buffer, and epoch-wise slow/meta updates are used to stabilize training, with zero inference-time model calls at deployment. The deployed artifact is a compact best_skill.md that runs against the unchanged target model.
Getting started
Install the package:
pip install skillopt
WebUI (optional):
pip install -e ".[webui]"
python -m skillopt_webui.app
The documentation and release notes provide the full installation, data preparation, training/eval commands, configuration, and framework internals.
Recent releases
Latest two releases:
- v0.2.0 SkillOpt v0.2.0 — SkillOpt-Sleep (2026-07-02) The headline feature is SkillOpt-Sleep: a nightly offline self-evolution engine behind a held-out validation gate; includes experimental multi-objective, replay, and dream-rollout controls; adds integration shells for Claude Code, Codex, Copilot, Devin; Windows robustness and hardened JSON parsing.
- v0.1.0 v0.1.0 (2026-06-02) — Initial PyPI Release
Install with
pip install skillopt.
Install command shown in the release notes:
pip install skillopt
Traction
Stars: 15586
Behind the repo
Not provided in the given facts.
Caveats
License: MIT Created: 2026-05-08 Last push: 2026-08-02