Google Research presents WikiSkill for evolving agents
Google Research’s WikiSkill framework co-evolves reusable agent skills with a persistent wiki that compiles execution traces into structured knowledge. Across five benchmarks and five models, it improves performance, supports cross-model skill transfer, and can help smaller models outperform larger ones without skills.
Persistent knowledge may be the missing layer between one-shot agent optimization and genuinely cumulative improvement.
- –Separating raw traces, curated patterns, and executable skills makes agent learning auditable and reusable
- –Validation gating accepts only skill changes that improve performance, while retaining rejected lessons in the wiki
- –WikiSkill improved average results across every tested model, with gains reaching 23.9 points for Qwen-3.6-27B
- –Cross-model transfer suggests skill discovery and skill execution are distinct capabilities
- –The approach is promising, but benchmark-specific workflows and full skill injection leave open questions about real-world retrieval and maintenance costs
DISCOVERED
1h ago
2026-08-28
PUBLISHED
1h ago
2026-08-28
RELEVANCE
AUTHOR
omarsar0