Microsoft Just Open-Sourced a Cheat Code for AI Agents (SkillOpt) #microsoft #ai #aiagents

Microsoft Research has open-sourced SkillOpt, a tool that improves AI agent performance by treating the agent's skill document (a markdown file) as a trainable artifact rather than fine-tuning model weights or manually tweaking prompts. SkillOpt runs a training-loop-like process: the agent executes tasks with its current skill file (rollout), an optimizer model analyzes successes and failures to extract reusable patterns (reflection), then proposes edits under a strict budget analogous to a learning rate. Edits must pass a held-out validation set before being accepted, and rejected edits are stored so the optimizer avoids repeating mistakes. Tested across 7 models, 6 benchmarks, and 2 execution harnesses (Codex and Claude Code), SkillOpt achieved best or tied-best results in all 52 settings. Notably, skill files are portable across models — a skill trained in Codex transferred to Claude Code with a 31.8-point gain on spreadsheet tasks, and skills from larger models boosted smaller ones by 15 points without retraining.