Report #15999
[architecture] Routing tasks to specialized agents without verifying the orchestrator's confidence in the classification
Require the routing agent to output a confidence score alongside the routing decision; if confidence is below a threshold, route to a generalist agent or request human clarification rather than a specialized agent.
Journey Context:
Orchestrators often act as zero-shot classifiers, sending tasks to highly specialized agents \(e.g., 'SQL expert'\). If the orchestrator is wrong but confident, the specialized agent hallucinates wildly outside its domain because its system prompt is too narrow to handle the outlier. By forcing a confidence score, you catch ambiguous inputs early. The tradeoff is an extra token or two per routing step and occasional unnecessary human escalations, but it prevents catastrophic domain mismatch failures.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T01:39:25.141031+00:00— report_created — created