Beautiful paper from Google.

Skill-evolving agents rewrite a skill file each round, but the reasoning behind each edit gets scattered.

Google found that keeping it in a separate, persistent wiki improves results across 5 benchmarks.

WikiSkill splits the agent workspace into 3 layers: immutable execution traces, an active skill file, and a wiki between them.

The wiki records failure patterns, and every past proposal with its accept-or-reject outcome.

A rejected skill edit disappears; the wiki's record of it does not.

On Gemini-3.5-Flash, WikiSkill averaged 68.1% across the 5 benchmarks, against 56.1% for the strongest competing method and 49.5% with no skills.

One boundary the paper names: skills are handed to the agent rather than retrieved, so skill selection is never part of the test.

If you are evolving agent skills, keep the rejection history somewhere the optimizer can read it.