WikiSkill – Google's Agent Skill Evolution Framework

1. What is WikiSkill

WikiSkill is an Agent skill evolution framework introduced by Google Research. It employs a three-tier architecture to separately store raw execution traces, structured knowledge, and executable skills, enabling an Agent's experience to be preserved as persistent knowledge and continuously evolve into reusable skills. In each iteration, the Wiki Maintainer analyzes the traces to update the knowledge base, while the Skill Proposer generates skill updates based on this knowledge. These updates are then evaluated on a validation set and either accepted or rolled back. Experiments show that a 9B parameter model combined with WikiSkill can outperform a 27B parameter base model, and the evolved skills can be transferred and reused across different models and model families.

Technical Positioning and Domain: WikiSkill belongs to the intersection of AI Agent autonomous evolution and skill learning, focusing on solving the core issue of how Agents can continuously learn from experience and accumulate reusable skills. This framework provides a new paradigm for the integration of reinforcement learning and LLM Agents, introducing a persistent knowledge layer between experience and skills, thereby enabling a knowledge compounding effect across iterations.

Development Background: This framework was developed by Google Research's Agent team. Based on their observation that LLM Agents often lack reusable domain-specific skills in open tasks—such as Excel formula processing or long document retrieval—despite their general reasoning capabilities—the team designed WikiSkill to endow Agents with the ability to continuously evolve autonomously.

Core Value: WikiSkill addresses the widespread "experience waste" problem in the skill learning process of LLM Agents. In traditional approaches, failed skill updates would roll back all previously accumulated knowledge, leading to inefficient learning. WikiSkill decouples the knowledge layer from the skill layer, ensuring continuous knowledge accumulation and safe skill rollbacks, thus achieving sustainable skill evolution and significantly reducing dependency on model parameter scale.

Technical Features: The framework introduces a unique three-tier knowledge separation architecture (Raw → Wiki → Skills), managing factual basis, structured knowledge, and executable skills separately. Its four-step evolution cycle (Execution → Knowledge Update → Skill Proposal → Validation Gating) ensures that all skill updates are data-driven and verifiable. Experiments have confirmed that skill discovery and execution are two distinct capabilities, and evolved skills can be transferred across different model families.

2. Key Features

• Three-Layer Knowledge Separation Architecture: The framework's core design lies in dividing an Agent's memory into three distinct layers: Raw Layer (raw trajectory layer), Wiki Layer (knowledge layer), and Skills Layer (skills layer). The Raw Layer permanently stores immutable execution trajectories as factual records; the Wiki Layer compiles trajectories into structured patterns and evolutionary logs, continuously accumulating knowledge without ever rolling back; the Skills Layer holds currently active executable skills, supporting conditional updates and rollbacks. Each layer operates independently and performs its own specific function without interference.

• Autonomous Skill Evolution Loop: The framework achieves continuous skill optimization through a four-step iterative loop. The Inference Agent executes tasks using the current skill set on the training set and generates trajectories; the Wiki Maintainer samples these trajectories for root-cause analysis and updates the knowledge catalog; the Skill Proposer reads from the knowledge base and trajectories using the ReAct method to propose new skill creation or patches; the Gating mechanism evaluates candidate skills on the validation set, accepting them only if performance metrics show a strict improvement, otherwise rolling back the skill. This entire loop requires no human intervention, enabling end-to-end autonomous evolution.

• Persistent Knowledge Compounding Accumulation: The Wiki Layer continuously accumulates structured knowledge across iterations. Even if a skill proposal is rejected by the validation gate, the pattern analysis and logs accumulated in the Wiki Layer are not lost. Subsequent iterations can reuse all historical insights, achieving knowledge compounding across iterations and avoiding redundant attempts from scratch.

• Cross-Model Skill Transfer: Experiments have shown that evolved skills can be reused across different models and even different model families. Researchers found that skill strategies discovered by other models can sometimes outperform those evolved by the model itself. This indicates that skill discovery and skill execution are two distinct capability dimensions, offering a new practical pathway for knowledge sharing between models.

• Validation Gate Filtering Mechanism: Candidate skills must undergo evaluation on a validation set. Only when performance metrics show a strict improvement are they accepted as new skill versions; otherwise, they automatically roll back to the previous state. This gating mechanism ensures reliable performance improvements with every skill update, while skill-impact.md fully records the differences, validation scores, and acceptance/rejection results of each proposal, forming an objective audit trail.

• Skill Traceability and Transparency: Each skill file achieves traceability through PURPOSE.md, clearly documenting the skill's purpose, origin, and design rationale. This design makes the skill library interpretable, allowing subsequent Proposers to understand the background and reasoning behind each skill, leading to more informed and rational update decisions.

• Avoiding Repeatedly Failed Approaches: The skill-impact.md file fully records all historical proposals, including diffs, validation scores, and acceptance/rejection results, forming an audit trail. Before submitting new proposals, subsequent Skill Proposers can refer to this file to avoid repeating approaches that have already been validated as failures, significantly improving the efficiency of the evolution process.

3. How to Use

1. Initialize the workspace: First, create a workspace containing three hierarchical directories: raw/, wiki/, and skills/. The raw/ directory is used to store raw execution traces, wiki/ is used to store structured knowledge files (including the patterns/ pattern directory, logs.md log file, and skill-impact.md impact records), and skills/ is used to store currently active skill files. The initialization phase starts with an empty skill set.

2. Execute training rollout: Configure the Inference Agent to perform tasks on the training set using the current skill set and generate traces. Each trace includes the complete execution process, intermediate states, and final results, and is written to the raw/ directory as factual basis for subsequent analysis. This phase does not require human intervention; all operations are automatically executed by the Agent.

3. Update the knowledge base: The Wiki Maintainer samples successful and failed traces from raw/, performs root-cause analysis, and creates or updates structured pattern files in wiki/patterns/. It also appends experience logs to logs.md and impact records to skill-impact.md. Knowledge generated in this phase is never rolled back, even if subsequent skill proposals are rejected.

4. Propose skill updates: The Skill Proposer reads the Wiki index, historical impact records, and raw traces, and proposes a single skill creation or patch using the ReAct method. Each proposal must be atomic—only targeting the creation or modification of a single skill, and only operating on a single skill file, not multiple files at the same time.

5. Validation and gating: Deploy the candidate skill on the validation set to run evaluations and compare the validation score with the baseline of the current skill. Only when the score strictly improves will the candidate skill be accepted and replace the existing skill in skills/; if the score does not improve or the candidate is rejected by the validation set, the skill will automatically roll back to a previous version, while the knowledge in the Wiki layer remains unaffected.

6. Iterative refinement: Repeat steps 2 through 5 until the validation performance converges or reaches the preset maximum number of iterations. After the iterations are complete, the skills/ directory will contain the final skill set that has evolved through multiple rounds, which can be directly used for deployment in the inference phase.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

6. Editor's Summary

The "three-tier knowledge separation" architecture proposed by WikiSkill represents innovative value not only in terms of storage decoupling at the engineering level, but more importantly, it redefines the paradigm of Agent skill learning. For a long time, failed Agent skill updates have often resulted in the loss of previously accumulated knowledge, a structural flaw that limited learning efficiency. WikiSkill completely separates "what is known" from "how to do it," enabling the knowledge layer to accumulate continuously and the skill layer to be safely rolled back, fundamentally addressing the issue of wasted experience. Experimental data also provides strong evidence—models with 9B parameters combined with WikiSkill can outperform 27B base models in task performance, and the discovery that skills can be transferred across different model families confirms that skill discovery and skill execution are indeed two distinct capabilities. This result has significant implications for the large-scale deployment of AI Agents, indicating that smaller models can perform tasks previously only achievable by large models through framework mechanisms, significantly reducing inference resources and deployment costs.

From a practical perspective, WikiSkill addresses the core need of enabling Agents to develop specialized skills that can sustainably evolve within specific domains (such as office automation, long document Q&A, and complex reasoning), rather than relying on the endless stacking of model parameters. Its four-step evolutionary cycle is well-designed, and the verification gating mechanism provides a reliable performance baseline for skill updates, while the audit records in skill-impact.md ensure transparency and traceability throughout the evolution process. This framework is suitable for AI researchers, Agent application developers, and engineering teams focused on model deployment efficiency. Currently, the framework is still in the research phase of the paper, and it is recommended to follow Google Research's subsequent open-source code and toolchain releases. Once mature, it will become a common infrastructure in the field of autonomous Agent evolution.

7. Application Scenarios

• Intelligent Office Automation: When handling complex Excel formula calculations, data cleaning, and cross-table associations in office tasks, the Agent can acquire specialized spreadsheet workflows such as "recalculate first, then read" through the evolutionary mechanism of WikiSkill. Once these skills are solidified, they become reusable persistent knowledge, continuously improving the efficiency of subsequent similar tasks.

• Long Document Intelligent Q&A: When performing multi-step retrieval navigation in hundreds of pages of PDF/Word documents, the Agent relies on evolved search skills to efficiently locate key paragraphs and avoid getting lost in long contexts. WikiSkill solidifies successful paths during retrieval into pattern files, making subsequent reasoning on similar document structures more goal-oriented.

• Mathematical Reasoning and Problem Solving: By solidifying the problem-solving experience of competition-level math problems into reusable strategies, the Agent continuously summarizes effective derivation paths and problem-type recognition patterns during its evolution, constantly improving problem-solving accuracy. Cross-iteration accumulation at the knowledge level ensures that each failed attempt becomes a basis for future reasoning.

• Web Information Retrieval: Optimizing multi-round search processes, the Agent evolves a standard operating procedure of "keyword expansion → multi-source cross-verification → information integration," enhancing recall and accuracy for complex queries. The retrieval patterns solidified in the Wiki layer allow the Agent to quickly adapt its strategy when facing search demands in different domains.

• Interactive Embodied Tasks: When performing multi-step physical operations in a virtual environment, the Agent improves task success rates by solidifying patterns such as "avoiding repetitive action loops" and "checking task completion status." The continuously evolving sequence of operations in the skill layer enables the Agent to stably handle changes in environmental states and abnormal branches.

8. FAQ

Q: What is the fundamental difference between WikiSkill and providing a skill library directly within an Agent?
A: Traditional skill libraries are static—all skills are fixed before deployment and do not optimize based on operational experience. The core of WikiSkill is "evolution"—each skill in the library undergoes a multi-round proposal, validation, acceptance, or rollback process, continuously refining optimal strategies from execution trajectories. The presence of the Wiki knowledge layer ensures that all evolutionary outcomes are accumulated and preserved, rather than starting over after each iteration.

Q: Does WikiSkill require a large amount of manually annotated data for initialization?
A: No. The framework only requires a training set containing task inputs and expected outputs, as well as a validation set for gate validation. Original trajectories are automatically generated by the Inference Agent using the current skill set. Root cause analysis, knowledge updates, and skill proposals are all automatically completed by the Wiki Maintainer and Skill Proposer, with no need for manual annotation of trajectories or skills. Human involvement is mainly limited to the design of the initial task set and the construction of the validation set.

Q: Will accumulated knowledge be lost if a skill is rejected during validation?
A: No. This is a core design principle of WikiSkill—decoupling knowledge from skills. When a skill proposal is rejected, only the skill in the Skills Layer is rolled back, while the pattern files, log records, and proposal history in the Wiki Layer are preserved and not deleted due to rejection. This knowledge continues to provide insights for future skill proposals in subsequent iterations, preventing the repeated attempt of previously failed approaches.

Q: Are the evolved skills in WikiSkill limited to specific models?
A: No, they are not limited to specific models. Experimental results have clearly shown that evolved skills can be transferred and reused across different models and model families. Research has even found that skills discovered by other models can sometimes outperform skills evolved by the model itself. This is closely related to the fact that skill discovery and skill execution represent two distinct capabilities—evolved skills are stored in text form and do not depend on the parameter structure of a specific model, giving them broad transferability and applicability.

9. Project Links

• Technical Report Paper: https://arxiv.org/pdf/2608.27454