Is predicting the right queue enough to evaluate a routing agent?

**Is predicting the right queue enough?**

No. Prediction identifies a proposed destination. A working agent must also establish whether it may assign the request, perform the permitted action and recognise whether the destination accepted it. These are separate capabilities.

**What can go wrong after a correct prediction?**

A professional services enquiry may belong to a restricted engagement queue. An intake agent can identify that queue but lack permission to assign work directly. The authorised next step may be a referral to an intake reviewer. Attempting a direct assignment is still a failure, even when the destination is correct.

**How should a task represent that distinction?**

A generated task could specify the agent's role, available tools and permitted transitions. Its expected outcome could distinguish a recommendation, a submitted referral and an accepted assignment. The agent's final response should accurately describe which state it reached.

**Why does this matter for training?**

If every correct label is rewarded as completed work, the agent learns to report success before the business process has changed. Evaluation should examine both the proposed destination and the action the agent was actually authorised to complete.