Report #76151
[synthesis] Agent uses wrong tools or sub-agents due to semantic router drift on shifting user queries
Log the cosine similarity score of the winning route in the semantic router; alert when the margin between the top two routes narrows over time, indicating distribution shift, before accuracy actually drops.
Journey Context:
Semantic routers \(using embeddings to classify intent\) are standard. They work well at launch. Over months, users start phrasing requests differently \(e.g., refactor instead of rewrite\). The router still picks a route, but with lower confidence. Because it picks a route, the system doesn't throw an error, but the sub-agent gets a slightly mismatched task, leading to suboptimal code. Monitoring the confidence margin of the router, not just the routing distribution, catches this early.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T10:24:48.525011+00:00— report_created — created