Escalation triggers

Definition

The three legitimate conditions under which an agent should hand a task to a human instead of continuing to work it autonomously.

Key points

• Three valid triggers:
  1. Explicit human request → escalate immediately, no investigation or attempt first (see Explicit-request-overrides rule).
  2. Policy is ambiguous or silent → don't improvise or invent a policy; escalate.
  3. Cannot make meaningful progress — repeated failures, missing data, or unresolvable errors (see Limits of automated retry).
• Complexity alone is NOT a trigger — a hard-but-tractable problem should still be worked, not escalated.
• Nuance: a frustrated customer → acknowledge, offer to resolve, and escalate only if they reiterate wanting a human — frustration by itself is a wrong signal.
• When escalating, always attach a Structured handoff summary; escalating with no context is itself a failure pattern.

Why it matters for the exam

• Domain 6 is unscored but frequently tested; scenarios plant a trigger (or a fake trigger) and ask whether to escalate.
• The distinction "meaningful progress impossible" vs "merely complex" is the classic distractor.

Common gotchas

• Complexity, sentiment, and confidence scores are traps — none is a valid trigger.
• Policy silence means escalate, not extrapolate.

See also

• Wrong escalation signals
• Explicit-request-overrides rule
• Structured handoff summary
• Limits of automated retry
• Human-in-the-loop calibration
• Minimal footprint principle

Sources

• netlight-study-guide
• anthropic-model-spec