Google WikiSkill: Persistent knowledge bases boost agent evolution
Saboo_Shubham_ · x · 2026-09-01
Google Research released the paper 'WikiSkill', proposing a framework that co-evolves agent skills with a persistent knowledge base (wiki) to compile experience. Key findings include:
- Wiki-driven evolution: Agents improve at tasks by keeping a wiki of everything they've learned.
- Small model wins: A small model with skills can outperform a model 3x its size without them.
- Knowledge is key: Gains disappear immediately if the wiki is deleted.
- Cross-model transfer: Evolved skills transfer effectively across models and families, and skills evolved by other models can outperform self-evolved ones.
More from coding & agent
- Harness, a command center for coding agents, hits #2 on Product Hunt — dee_hw · 2026-10-16
- nanobot: a 4,000-line rewrite of openclaw, 99% smaller — mdancho84 · 2026-10-16
- Open-source Free WebAccess routes agent web search across vendors for $0 with auto-fallback — yangyi · 2026-10-16
- State of Agentic Coding podcast hosts Geoffrey Huntley on code mode and self-modifying harnesses — mitsuhiko · 2026-10-16
- The Road to Self-Improving Software: Building a Closed-Loop Agent Factory — _AustinCalvert_ · 2026-10-16
- Screenshots show "talk-to-build" AI app tool with one-click open-source publishing to GitHub — yangyi · 2026-10-16