Report #35839
[synthesis] Multi-agent router sends a specialized task to the wrong agent, which hallucinates a solution outside its expertise
Keep the router's taxonomy strictly orthogonal, and give each specialized agent a strict 'rejection' capability to bounce tasks back to the router if the task falls outside its defined scope.
Journey Context:
In multi-agent systems, developers want a central router to delegate. But if agent capabilities overlap \(e.g., 'code writer' vs 'code reviewer'\), the router might send a writing task to the reviewer. The reviewer, trained to be compliant, will try to write the code badly instead of rejecting it. The synthesis is that agent compliance is a liability in routing; agents must be explicitly trained/prompted to reject mismatched tasks to prevent silent degradation of service. Without rejection, the system lacks a feedback mechanism for routing errors.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T14:38:06.760273+00:00— report_created — created