A persistent wiki layer between traces and skills is the missing piece in agent evolution

The core insight in WikiSkill is a separation that prior skill-evolution frameworks collapse: raw execution traces, accumulated knowledge, and executable procedural skills are three distinct things, and conflating them limits how much an agent can learn across iterations. Previous methods like EvoSkill, Trace2Skill, and SkillOpt all run some version of the same loop—roll out, analyze traces, patch a skill document, gate on validation—but none maintain a separate, persistent knowledge representation that compounds across those iterations. WikiSkill adds exactly that: a Wiki Layer sitting between immutable execution logs and the active skill files, populated by a dedicated Wiki Maintainer agent that extracts failure patterns and successful strategies into individual markdown pages, logs every proposal and its acceptance outcome, and never gets rolled back even when a skill update is rejected.

The empirical case for this design is substantial. Across five benchmarks and five models from three families, WikiSkill beats the strongest competing method for every model tested, with margins ranging from 3.3 points for the smallest Qwen model up to 12.0 points for Gemini-3.5-Flash. The scaling interaction is the most interesting finding: within the Qwen family, average improvement from WikiSkill grows from 12.3 points at 4B parameters to 17.5 at 9B and 23.9 at 27B. The SpreadsheetBench numbers make this concrete—gains of 6.5, 9.3, and 40.9 points respectively across those three sizes. At the same time, a 9B model with WikiSkill-evolved skills reaches 47.4% average accuracy against 39.4% for a 27B model running without any skills.

The cross-model transfer results complicate the picture in a productive way. Skills evolved by one model frequently outperform self-evolved skills when transferred to another. On ALFWorld, Qwen-3.5-9B reaches 70.2% using a skill evolved by the 27B model, versus 63.4% with its own. Transfer can also flow upward from smaller to larger models: 4B-evolved skills improve Gemma-4-31B to 73.1% on LiveMath. But negative transfer is real. Qwen-3.5-4B SpreadSheet skills drop Gemini-3.5-Flash from 50.5% to 18.1%, because the smaller model encodes low-level workarounds—single-line Python commands, string-conversion rules—that actively constrain stronger models and exhaust their interaction budgets with redundant tool calls. This distinguishes two capabilities that self-evolution normally conflates: discovering useful procedural knowledge and executing it reliably.

The ablation is decisive on the central claim. Giving the Skill Proposer access to the persistent wiki raises average benchmark performance from 48.7% to 63.7% on Gemini-3.5-Flash, with LiveMath jumping from 51.3% to 72.6%. Counterintuitively, giving the Inference Agent wiki access during training rollouts hurts performance—the hypothesis being that agents solving tasks directly from wiki patterns produce trajectories that are less informative for skill refinement.

One genuine limitation: skills are currently injected wholesale into the agent's system prompt rather than retrieved selectively, which sidesteps the retrieval problem entirely. The wiki also has no automated pruning mechanism, which will matter as evolution runs grow longer. These are acknowledged gaps, not oversights.