Microsoft open-sources SkillOpt: train agent skills like neural networks

aigclink · x · 2026-08-31

Microsoft has open-sourced SkillOpt, which applies neural-network-style training to skill documents: it only optimizes the skill doc, not the underlying model weights, producing a compact bestskill.md. Unlike hand-written or self-modified skills whose improvement is unverifiable, SkillOpt offers an evaluable training loop. On GPT-5.5 in direct conversation, it reportedly gains 23.5 points over the no-skill baseline.

Its flagship SkillOpt-Sleep feature evolves skills offline overnight: it harvests daytime sessions, mines recurring tasks, replays exercises, and merges skills that pass tests into the guide. Deployment adds zero extra inference calls and no added latency.

Related event: Microsoft Open-Sources SkillOpt: Training Agent Skills Without Touching Weights(2 posts)→

Original post →

More from coding & agent

coding & agent channel →