SkillOpt: Turning AI Agent Skills into Trainable Parameters

Microsoft Research's paper 'SkillOpt: Executive Strategy for Self-Evolving Agent Skills' reframes agent skill editing as a training process. SkillOpt treats the skill file as a trainable parameter outside a frozen target model, using a forward-backward-update cycle in text space. In the forward pass, the frozen model executes training tasks; the backward pass uses an optimizer model to analyze trajectories; the update step proposes bounded edits gated by a validation check. Rejected edits are stored for negative feedback, and a slow/meta update consolidates longer-horizon lessons. Evaluated on six benchmarks (SearchQA, SpreadsheetBench, OfficeQA, DocVQA, LiveMathematicianBench, ALFWorld) with seven target models (from GPT-5.5 to Qwen3.5-4B) and three execution modes (direct chat, Codex, Claude Code), SkillOpt achieved best or tied-best results in all 52 cells. For GPT-5.5 in direct chat, the six-benchmark average rose from 58.8 to 82.3 (+23.5 points). Gains were largest on procedural benchmarks: SpreadsheetBench from 41.8 to 80.7, OfficeQA from 33.1 to 72.1, and LiveMathematicianBench from 37.6 to 66.9. Optimized skills transferred across model scales, harnesses, and related tasks; for example, a spreadsheet skill trained in Codex transferred to Claude Code improved the no-skill baseline from 22.1 to 81.8 (+59.7). Final skill files remained compact (median ~920 tokens) with only 1-4 accepted edits. SkillOpt suggests that training can be applied to procedural knowledge outside model weights, offering a controllable, auditable adaptation layer for agentic workflows.