Skill Routing in Practice: Self-Evolving Descriptions from Real Requests
When an agent picks the wrong Skill, fix it with local first-route replay plus a loop of attribution, rewrite direction, description edit, regression validation and recorded lessons — 31 of 34 dev requests routed correctly, 91.18%.
The naive fix for a routing error is to hand the failed request to a model, let it rewrite the description, and re-run the whole batch. That produces a new version but answers nothing: why this change, does it target the real conflict, what should the next round do. This practice splits the problem in two — local first-route replay to shorten the feedback cycle, and an iteration loop organized as attribution and optimization direction, concrete description rewrite, regression validation, then recorded lessons reused next round. The running example is two adjacent capabilities that both query data but differ in delivery: one returns the result inside the conversation, the other pushes a notification to an external channel. Writing out the delivery purpose behind the shared steps explains the fix better than adding more query keywords. The selected candidate routed 31 of 34 development requests correctly — 91.18% — a dev-set best pick, not a claim about unseen requests or full online task success. Description growth and feedback cost are tracked as explicit constraints.
The full text of this article is currently available in Chinese only. Read the original