Google's WikiSkill Boosts Agent Accuracy by 15 Points

Google Research and Virginia Tech have published an August 2026 paper detailing WikiSkill, a framework designed to evolve AI agent instructions by compiling experience into persistent knowledge. Instead of manually updating instruction files or relying on evolution loops that discard failed attempts, WikiSkill maintains a private, persistent notebook of past failures and workarounds. This asymmetric memory structure prevents agents from wasting computational resources on rediscovering dead ends. The workspace is split into three layers: a raw folder for execution traces, a wiki folder that compiles pattern pages and an append-only ledger called skill-impact.md, and a skills folder for the active instruction files.

During training, an actor agent attempts tasks using current skills, a maintainer agent patches pattern pages from a sample of up to five failures and three successes, and a proposer agent suggests skill changes. On Gemini-3.5-Flash, giving the proposer agent access to this persistent wiki boosted average accuracy from 48.7 to 63.7. However, letting the actor agent peek at the same wiki dropped performance back to 60.9. The framework also improved LiveMath scores from 51.3 to 72.6 and Spreadsheet scores from 49.9 to 76.6. On the ALFWorld benchmark, using Qwen-3.6-27B, the system successfully refined a rejected "goal-directed-action" skill into a concrete rule by keeping the failed attempt in its ledger.

For practitioners, WikiSkill offers a structured way to manage agent evolution without manual instruction rotting. The framework's SkillOpt ledger prices this evolution at 0.6 million to 46.4 million training tokens per accuracy point for artifacts of 379 to 1,995 tokens, from a median of 2.5 accepted edits. Notably, a 9B model equipped with these skills scored 47.4, outscoring a larger 27B model. However, skill transfer requires caution; while transferred skills beat self-evolved ones 70.2 to 63.4 on ALFWorld, a 4B spreadsheet skill crashed a stronger model's performance from 50.5 to 18.1, highlighting the necessity of the skill-impact.md promotion gate. Only 39 to 52 percent of accepted updates land in the first two rounds, with 28 percent arriving late. The authors have provided an appendix scaffold that turns the entire loop into an hour-long runnable test.