3 · Tools, Agents & MCP › Phase 4 › Lesson 6 of 6

When agents should give up

What this lesson covers

An agent that never gives up is not a virtue; it’s a $400 invoice. This lesson is about the design decision most teams put off until an incident forces it.

The outline

1. The four “stuck” conditions. Same tool called with same args twice, no progress on a plan step, budget exceeded, low-confidence answer.
2. Detecting each from telemetry. What to log so you can spot stuck agents before your bill does.
3. The handoff. What state you hand to a human, in what shape, and via what channel.
4. Confidence, honestly. How to get calibrated “I don’t know” signals from a model that will happily invent an answer instead.
5. The retry-vs-hand-off decision. When the right response is to try again with a different plan, and when it’s to stop.
6. Metrics that reward giving up. Why measuring “task completion rate” punishes correct handoffs, and what to measure instead.

Coming soon

In outline.