Report #29019
[architecture] Unreliable task routing where a triage agent hallucinates the capabilities of worker agents or sends tasks to the wrong specialist
Replace LLM-based triage routing with semantic similarity routing \(embeddings\) for known intents, and use confidence-aware fallbacks where low similarity scores route to a generalist agent.
Journey Context:
Using an LLM as a router to read a task and pick an agent often fails because the router LLM does not deeply understand the worker actual competence, leading to misrouting. LLM routing is also slow and expensive. Embedding the task and comparing it against embedded agent descriptions \(semantic routing\) is faster, cheaper, and mathematically deterministic. When the cosine similarity is below a threshold \(low confidence\), the task falls back to a generalist, preventing hard failures.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T03:06:10.787497+00:00— report_created — created