The problem WikiSkill attacks is the one every agent builder hits: an agent solves something hard, and the insight evaporates when the run ends. Recent work discovers reusable skills from agent experience, but as the paper puts it, the insights that guide skill development stay "scattered across optimization histories," so nothing compounds. Google Research's answer, in a paper posted this week as arXiv 2608.27454 by Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan and Tu Vu, is to give the agent a wiki.
The design is three layers that are deliberately not the same thing. The raw layer keeps complete execution logs. The wiki layer holds distilled insights about what failed, what worked and why, and is never reset. The skill layer holds procedural instructions, and can be reset when performance degrades. Each iteration runs execution, analysis, skill proposal, then validation through a gate before anything is written down. Nothing here touches model weights: the learning lives in text that the same model reads next time.
The numbers, as reported by The Decoder, cover five task families, mathematical reasoning, web search, spreadsheet manipulation, document question answering and interactive environments, with models from Qwen at 4B to 27B, Gemma and Gemini. Gemini 3.5 Flash goes from 49.5 to 68.1 percent on average; Qwen-27B from 39.4 to 63.3; the largest single gains are LiveMath at 33.0 to 72.6 and spreadsheet tasks at 50.5 to 76.6. The reported pattern is that larger models extract more from evolved skills, while smaller models can approach larger models' performance by borrowing them.
The limitations are worth as much as the headline. Gains vary sharply by task type, long-document contexts barely move, and skills transferred between models sometimes hurt. Still, the direction is the interesting part for anyone running agents in production: this is not a bigger context window or a better retriever, it is a curator, an explicit step that decides what an experience meant before storing it. If the encoding step is what separates a library from a librarian, this paper is a serious attempt to write the librarian down.