Core concept: An escalation trigger must be a fact the system can check, not a feeling it has to judge. Anthropic's eval guidance sets the bar: "A good task is one where two domain experts would independently reach the same pass/fail verdict." Write triggers that clear that bar.

Why A and C are correct:

C: an explicit request for a person is a visible event in the transcript. Two reviewers would always agree it happened.
A: a policy gap is visible too. The policy lookup returns no matching rule, so the agent has nothing to apply. Continuing means inventing policy.
"Building effective agents" says agents "pause for human feedback at checkpoints or when encountering blockers". No applicable rule is a blocker.
The support guide grades "escalation efficiency", tracking cases "that should have been escalated but weren't". Crisp triggers are what close that gap.

Why the others are wrong:

B: the number looks objective, but it is the model's own self-report. It is not calibrated and nothing verifies it. A 0.61 and a 0.59 are not different facts.
D: sentiment is a judgment call, and frustration alone does not mean the agent cannot help. Acknowledge it, offer the quick fix, and escalate when the customer actually asks.

Remember: Escalate on events you can point at, never on scores or moods.

Question 163