MicrosoftSkillOpt trains an agent's natural-language skill document instead of the model's weights. The target model, backend, and harness stay fixed, and the skill document that guides evidence gathering, tool use, verification, and output formatting is treated as trainable state, optimized with epochs, minibatches, a learning rate, and validation gates. The output is a compact best_skill.md file (typically 300 to 2,000 tokens) that runs against the unchanged model and adds no inference-time model calls.
Each training step runs the frozen agent on scored batches and records messages, tool calls, verifier feedback, and scores. A separate optimizer model reflects on failures and successes separately and proposes bounded add, delete, and replace edits to the skill, and a candidate is kept only when it improves a held-out validation score. An edit budget acts as a textual learning rate so broad rewrites cannot overwrite useful rules, and a rejected-edit buffer plus an epoch-wise slow and meta update stabilize training.
Across six benchmarks, seven target models, and three execution harnesses (direct chat, Codex CLI, and Claude Code CLI), SkillOpt is best or tied-best in all 52 evaluated cells, and optimized skills transfer across model scales, between Codex and Claude Code, and to nearby benchmarks. On GPT-5.5 it raises average no-skill accuracy as follows.
| Harness | Accuracy gain (points) |
|---|---|
| Direct chat | +23.5 |
| Codex agentic loop | +24.8 |
| Claude Code | +19.1 |
SkillOpt-Sleep is a nightly offline companion for local coding agents (Claude Code, Codex, Copilot) that reviews past sessions, replays recurring tasks, and consolidates validated skills behind a held-out gate. The repository also includes integration shells for Claude Code, Codex, Copilot, and Devin, and an OpenClaw reference adaptation.
Features
- Skill training loop: rollout, reflect, aggregate, select, update, and evaluate, applied to a single skill document
- Validation gating: a candidate edit is accepted only when it improves a held-out validation score
- Textual learning rate: an edit budget bounds how far the skill can move in one update
- Deployable artifact: a compact best_skill.md that runs against the unchanged target model
- Backends: OpenAI, Azure, Claude, Qwen, MiniMax, Copilot, any OpenAI-compatible endpoint, and execution backends for Codex, Claude Code, Cursor, and Copilot
- Built-in benchmarks: six benchmarks, with a documented contract for adding new ones
- SkillOpt-Sleep: a nightly engine that harvests, mines, replays, and consolidates skills from past coding-agent sessions
- Agent integrations: integration shells for Claude Code, Codex, Copilot, and Devin, plus an OpenClaw reference adaptation
- WebUI dashboard: an optional Gradio dashboard for monitoring training
