Microsoft SkillOpt Boosts AI Agent Performance by Optimizing Skill Documents

Microsoft Research has introduced SkillOpt, a system that improves AI agent performance by automatically refining text-based skill documents rather than modifying any underlying model weights. A separate optimizer model analyzes agent behavior trajectories, identifies errors, and proposes bounded edits to markdown skill files within a controlled budget per iteration. Proposed changes are validated against held-out test sets and only accepted if they demonstrably improve scores, while a rejected-edit buffer prevents the optimizer from repeating ineffective suggestions. In testing across 7 models, 6 benchmarks, and 3 execution environments, SkillOpt achieved best or tied-best results in all 52 evaluated cells, with GPT-5.5 gaining up to 24.8 percentage points on certain tasks. The approach offers a cost-effective alternative to fine-tuning, requiring no changes to model architecture, weights, or inference costs.