Google's WikiSkill Framework Lets Agents Turn Experience into Reusable Skills
xiaohu · x · 2026-09-01
Google Research introduced the WikiSkill framework, designed to enable AI Agents to accumulate experiences and lessons from tasks into a persistent knowledge base (like a Wiki), thereby continuously evolving more powerful executable skills.
Core Pain Points
Existing Agent skill evolution methods primarily optimize the "current Skill" without maintaining a separate, growing knowledge base. This leads to:
- Experience is hard to reuse across different rounds.
- The value of failed modifications is easily lost.
- Skills become bloated as they bear too much responsibility (saving both experience and execution rules).
Three-Layer Knowledge Architecture
WikiSkill addresses the knowledge gap with a layered workspace:
- Raw Layer: Saves complete execution trajectories as immutable evidence.
- Wiki Layer: Saves understanding by organizing common patterns from trajectories into independent pages.
- Skills Layer: Saves action rules for the executing Agent, which remain traceable to underlying evidence.
Performance
WikiSkill outperforms existing baselines on average across five test models. A 9B model with WikiSkill skills even surpasses the performance of a 27B model without skills.
More from coding & agent
- Inside OpenAI's Codex Comeback: From Failed Launch to the App That Swallowed ChatGPT — every · 2026-10-16
- Red Teaming Production AI Agents: What Breaks Under Targeted Attacks — njyx · 2026-10-16
- AI models ran transport companies in OpenTTD for 10 years: Opus won with £45.8M — rachittshah · 2026-10-16
- One Prompt, 629 Lines: IQuest-Q1 Builds a Deep-Sea Submarine Game in 2 Minutes — Div_pradeep · 2026-10-16
- Swapping RLM sub-calls for Jev: half the cost, 6x faster at 1M tokens — lateinteraction · 2026-10-16
- Developer trains his own PII detector, dissatisfied with existing solutions — HankYeomans · 2026-10-16