WikiSkill - Google's Agent Skill Evolution Framework

What is WikiSkill

WikiSkill is an Agent skill evolution framework launched by Google Research. Through a three-layer architecture—the Raw Layer that stores raw trajectories, the Wiki Layer that continuously accumulates structured knowledge, and the Skills Layer that holds executable skills—the framework allows Agent experience to settle into persistent knowledge, from which reusable skills then grow. In each iteration, the Wiki Maintainer analyzes trajectories to update the knowledge base, the Skill Proposer proposes skill updates based on the knowledge, and after validation they are accepted or rolled back. Experiments show that a 9B model plus WikiSkill can surpass a 27B bare model, and skills can transfer across models.

WikiSkill's Main Features

Three-layer knowledge separation: Stores raw trajectories, structured knowledge, and executable skills in three separate layers, each with its own role.
Autonomous skill evolution: Through iterative loops, Agent skills continuously grow and optimize from experience.
Persistent knowledge accumulation: The Wiki layer continuously accumulates across iterations, and rejected skills do not lose already accumulated knowledge.
Cross-model transfer: Evolved skills can be reused across different models and even different model families.
Validation-gated filtering: Candidate skills are evaluated on a validation set, and only updates that improve performance are accepted.

WikiSkill's Technical Principles

Three-layer architecture: The Raw Layer permanently stores immutable execution trajectories as factual evidence; the Wiki Layer compiles trajectories into structured patterns and evolution logs, continuously accumulating and never rolling back; the Skills Layer holds the currently effective executable skills, supports conditional updates and rollback, and implements skill traceability through PURPOSE.md.
Four-step evolution loop: The Inference Agent executes on the training set using the current skills and produces trajectories; the Wiki Maintainer samples trajectories for root cause analysis and updates the Wiki's pattern catalog and logs; the Skill Proposer uses ReAct to read the Wiki and trajectories and proposes a single skill creation or patch; the Gating mechanism evaluates candidate skills on the validation set, accepting them if the score improves, otherwise rolling back the skill (the Wiki is unaffected).
Knowledge persistence design: The Wiki layer records the diff, validation score, and accept/reject result of each proposal through skill-impact.md, forming an objective audit trail that enables subsequent Proposers to avoid repeating failed approaches, achieving knowledge compounding across iterations.
Separation of skill discovery and execution: Experiments show that skill discovery (distilling strategies from experience) and skill execution (applying strategies at inference time) are two different capabilities. Skills evolved by WikiSkill can transfer across models, and skills evolved by other models can even outperform skills evolved by the model itself.

How to Use WikiSkill

Initialize the workspace: Create the raw/、wiki/、skills/ three-layer directories, starting from an empty skill set.
Run training rollout: The Inference Agent runs on the training set using the current skills, producing execution trajectories written to raw/。
Update the knowledge base: The Wiki Maintainer samples successful/failed trajectories for root cause analysis, creates or updates pattern files in wiki/patterns/ , and appends logs.md and skill-impact.md。
Propose skill updates: The Skill Proposer reads the Wiki index, historical impact records, and raw trajectories, and uses ReAct to propose a single skill creation or patch.
Validation and gating: Candidate skills are evaluated on the validation set; if the score improves, they are accepted as the new skills/。
Loop iteration: Repeat the above steps until validation performance converges or a preset number of iterations is reached, ultimately obtaining the evolved skill set for inference.

WikiSkill's core advantages

Persistent knowledge compounding: The Wiki layer continuously accumulates across iterations; when skills are rejected, knowledge is not lost, and subsequent updates can reuse all historical insights.
Knowledge-skill decoupling: Separating "what to know" from "how to do it"; skills can be rolled back while knowledge persists, avoiding experience being lost along with skill patches.
Small model beats large model: Qwen-3.5-9B plus WikiSkill can surpass the bare Qwen-3.6-27B model, using knowledge to compensate for the scale gap.
Cross-model transfer and reuse: Evolved skills can transfer across model families, and strategies discovered by other models are sometimes better than the skills evolved by the model itself.
Avoid repeating pitfalls: skill-impact.md fully records the proposal history and rejection reasons, so the Proposer no longer repeats failed proposals.

WikiSkill's project address

arXiv technical paper：https://arxiv.org/pdf/2608.27454

Comparison of WikiSkill with similar competitors

Comparison dimension	WikiSkill (Google Research)	SkillOpt (Microsoft Research)
Core idea	Add a persistent knowledge layer (Wiki) between experience and skills; knowledge continuously accumulates, and skills grow from knowledge	Treat skill documents as trainable external state, iteratively updated with deep-learning-style optimizers
Knowledge management	Three-layer architecture separation: Raw → Wiki → Skills; Wiki never rolls back, supporting cross-iteration knowledge compounding	A single skill document serves as the training state, with no independent persistent knowledge layer, relying on a rejection edit buffer and meta-guidance
Evolution mechanism	The Wiki Maintainer performs root cause analysis and updates the structured knowledge base, and the Skill Proposer proposes updates from the knowledge	Six-stage ReflACT pipeline: Rollout → Reflect → Aggregate → Select → Update → Evaluate
Editing strategy	Incremental patch editing (create/modify skills); each atomic proposal targets only a single skill	Structured add/delete/replace edits, introducing an edit budget as the text learning rate
Validation gating	If the validation set strictly improves, accept; otherwise roll back the skill (Wiki is unaffected)	If the validation set strictly improves, accept; rejected edits are stored in a buffer as negative feedback
Cross-model transfer	Explicitly verified: skills can transfer across model families, and skills evolved by other models outperform self-evolution	Supports cross-platform transfer (Codex ↔ Claude Code) without retraining

WikiSkill's application scenarios

Intelligent office automation: Handling complex Excel formula calculations, data cleaning, and cross-sheet associations; the Agent learns spreadsheet-specific workflows such as "recalculate first, then read" through evolution.
Long-document intelligent question answering: Performing multi-step retrieval navigation in hundreds of pages of PDF/Word documents; the evolved search skills can guide the Agent to efficiently locate key passages and avoid getting lost in long contexts.
Mathematical reasoning and problem solving: Distill the problem-solving experience of competition-level mathematics problems into reusable strategies, continuously improving solution accuracy.
Web information retrieval: Optimize the multi-round search process, evolving a standard operating procedure of "keyword expansion → multi-source cross-validation → information integration" to improve recall and accuracy for complex queries.
Interactive embodied tasks: Execute multi-step physical operations in a virtual environment, and improve task success rates by distilling patterns such as "avoid repetitive action loops" and "check task completion status".