RASO: borrow skills from other harnesses—adapt them, then validation-gate the update
Sep 29 Chu/Lee/Park et al.: retrieve GitSkills sections, Cross-Harness Adaptation into target tools; RASI init without rollouts + RASU feedback updates; beats SkillOpt/WikiSkill/GEPA on OfficeQA/Spreadsheet/ALFWorld/WebShop.
Source: Retrieval-Augmented Skill Optimization via Cross-Harness Adaptation
Jaewon Chu, Ji Soo Lee, Jihwan Park, and collaborators (Korea University / KAIST / Meta AI) note that millions of public agent skills encode procedural knowledge across tasks and harnesses—yet skill optimizers mostly ignore that corpus and burn rollouts refining text from scratch. RASO (Retrieval-Augmented Skill Optimization) treats an external skill library (GitSkills) as prior knowledge. RASI builds an initial skill from task + harness descriptions alone: generate requirement queries, BM25-retrieve top-K skill sections, then Cross-Harness Adaptation rewrites them into lessons that use only the target harness’s objects, commands, and units. RASU continues after failures: textual gradients plus failure-driven retrieval, same adaptation, then a skill updater—candidate kept only if validation improves.
On OfficeQA, SpreadsheetBench, ALFWorld, and WebShop with GPT-5.6-Luna and Qwen-3.5-9B, RASI beats retrieval-free init and naive SkillRouter reuse (on SpreadsheetBench 92.9% of retrieved skills come from a different harness). Full RASO tops TextGrad, GEPA, SkillOpt, and WikiSkill—e.g. GPT-5.6-Luna Spreadsheet 63.33 vs SkillOpt 57.02, OfficeQA 49.03 vs 45.54. Ablations: adaptation itself is load-bearing (RASI without it drops Spreadsheet 49.17→41.43); init and update each help, and together help most. At matched rollout budgets RASO is often cheapest among the optimizers (OfficeQA $0.84 vs TextGrad $1.45).
Why it matters for code review
Self-improving factories that rewrite overnight skills from their own traces alone are leaving the public skill corpus on the table—and pasting foreign skills without adaptation can hurt. RASO’s receipt matches continuous harness evolution under load: retrieve procedural priors, strip the wrong tools, and refuse skill edits that fail validation. Score skill libraries by cross-harness transfer after adaptation, not by how many skill.md files you ingested. FOMO for unchecked skill evolution from rollouts alone just overfits your training split while ignoring lessons already written for someone else’s harness.