Verified project record

microsoft/SkillOpt

skillopt treats the skill document as trainable state for frozen LLM agents, applying bounded text edits and validation gates to improve accuracy without touching model weights. The full training loop operates through rollout, reflect, aggregate, select, update, and evaluate phases, and outputs a compact best_skill.md artifact that adds zero inference-time model calls at deployment.

Project overview

By optimizing the skill document text rather than model weights, skillopt applies deep-learning-style training discipline to agent skills; edits are accepted only when they strictly improve a held-out validation score, and the resulting artifact runs against the unchanged target model with no additional inference-time cost.

Deployment: Refer to project documentation
License: MIT

Best for

AI engineers who need to optimize agent skill documents for frozen LLMs through a structured training loop with validation gates.
Teams seeking a deployable skill artifact that adds zero inference-time model calls against an unchanged target model.

Key capabilities

Full training loop for optimizing agent skills via rollout, reflect, aggregate, select, update, and evaluate phases.
Outputs a compact best_skill.md file that runs against the unchanged target model adding zero inference-time model calls.
A nightly offline self-evolution engine (harvest, mine, replay, consolidate) shipped as the skillopt-sleep CLI.
An optional monitoring dashboard launched via python -m skillopt_webui.app.

Limitations and risks

Integration plugin and MCP files for Claude Code, Codex, Copilot, and Devin live in the repository rather than the PyPI wheel.
Setup requires configuring external LLM backend APIs to run training and evaluation commands.

Getting started

Install with pip install skillopt, configure an external LLM backend API, run training or evaluation commands, and optionally launch the WebUI dashboard.

Evidence and sources

Open original GitHub project