Report #2981
[architecture] Agents fail to select the correct tool because multiple tools or agents have overlapping or ambiguous descriptions
Ensure tool and agent descriptions are mutually exclusive and collectively exhaustive \(MECE\) in the router's context, testing embeddings/similarity of descriptions to catch overlaps before runtime.
Journey Context:
LLMs select tools/agents based on semantic similarity between the prompt and the description. If 'Refund Agent' and 'Billing Agent' both mention 'handling payment issues', the router will randomly oscillate between them. Developers often write descriptions in isolation. The fix is to write descriptions as a MECE set and programmatically check their embedding distances. If two descriptions are too similar, merge the tools/agents or strictly partition their keywords. The tradeoff is that descriptions become less 'natural' and more rigid, but routing accuracy improves dramatically.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T14:43:05.051760+00:00— report_created — created