microsoft/SkillOpt

What it solves

SkillOpt addresses the problem of inconsistent and hand-crafted agent skills. Most AI agent skills are either written by hand or generated in a single pass by an LLM, which often fails to reliably improve over time based on feedback. SkillOpt provides a systematic, reproducible way to optimize these skills without needing to modify the underlying model's weights.

How it works

SkillOpt treats a skill document (a Markdown file) as the "trainable state" of a frozen LLM. It uses a separate optimizer model to iteratively improve this document through a training loop: rollout, reflection, aggregation, selection, and update.

Key mechanisms include:

• Validation Gates: Candidate edits to the skill document are only accepted if they strictly improve a held-out validation score.
• Optimization Parameters: It applies deep-learning concepts like epochs, batch sizes, and learning rates to the textual edits.
• SkillOpt-Sleep: An offline engine that harvests past sessions, replays tasks, and consolidates validated skills during "sleep" periods.
• Zero Overhead: Because the final output is just a compact Markdown file (best_skill.md), there are no additional model calls required during actual deployment.

Who it’s for

It is designed for developers building AI agents (such as those using Claude Code, Codex, or Copilot) who want to automatically upgrade their agent's capabilities and performance across different benchmarks and harnesses.

Highlights

• Model Agnostic: Works across various backends including OpenAI, Azure, Claude, Qwen, and MiniMax.
• High Performance: Demonstrated significant accuracy lifts (up to +24.8 points) on GPT-5.5 across multiple benchmarks.
• Transferable: Optimized skills can transfer across different model scales and execution harnesses.
• Offline Evolution: Includes a nightly self-evolution engine for local coding agents.