SkillOpt is a text-space optimizer developed by Microsoft that enables training of reusable natural-language skills for frozen large language model agents without modifying model weights. The system treats skill documents as trainable state and applies neural-network-style training discipline—including epochs, batch sizes, learning rates, and validation gates—entirely within the text domain. This approach allows skill improvement through trajectory-driven edits while maintaining deployment simplicity and zero inference-time overhead.
The core mechanism operates by having a separate optimizer model convert scored agent rollouts into bounded text edits on a single skill document. These edits take the form of additions, deletions, or replacements within the skill text. In the default configuration, candidate edits are accepted only when they strictly improve performance on a held-out validation set, ensuring stable and reproducible skill training. The system incorporates a textual learning-rate budget, a rejected-edit buffer, and epoch-wise slow and meta updates to maintain training stability.
SkillOpt's deployed artifact is a compact best_skill.md file, typically containing 300 to 2,000 tokens, that runs against the unchanged target model without requiring any modifications to the underlying LLM. The repository demonstrates strong empirical results across six benchmarks, seven target models, and three execution harnesses including direct chat, Codex CLI, and Claude Code CLI. Performance improvements are substantial: on GPT-5.5, SkillOpt lifts average no-skill accuracy by 23.5 points in direct chat, 24.8 points inside the Codex agentic loop, and 19.1 points inside Claude Code. Notably, optimized skill artifacts transfer across model scales, between different execution harnesses, and to nearby benchmarks without requiring further optimization.
The repository is written in Python and provides comprehensive documentation including installation guides, training and evaluation commands, configuration details, and framework internals. A versioned documentation system is maintained alongside a technical blog with engineering analysis. The project includes a WebUI dashboard for monitoring training progress and supports multiple backend integrations including OpenAI, Azure, Claude, Qwen, and MiniMax. The system also features SkillOpt-Sleep, a nightly offline self-evolution engine that harvests, mines, replays, and consolidates skills behind a validation gate.
SkillOpt v0.2.0 was released on PyPI with the SkillOpt-Sleep feature as a CLI tool, experimental multi-objective and replay controls, and integration shells for Claude Code, Codex, Copilot, and Devin. The framework is designed for extensibility, allowing users to add new backends by implementing a chat or execution target interface and new benchmarks through a standardized adapter pattern. The repository has received significant media coverage from Microsoft Research, VentureBeat, and other technology publications, and has been integrated by projects including gbrain, gbrain-evals, and darwin-skill. The system represents a novel approach to agent skill optimization that maintains reproducibility and stability while achieving consistent improvements across diverse evaluation scenarios.
