Learn · Context management & reliability · CCAR-F task 5.2

Escalation and ambiguity

Good agents know when not to act alone. Write escalation triggers as rules, gate high-stakes actions in code, and hand off with context.

Key ideas

• Write escalation triggers as explicit rules: explicit request for a human, a policy gap, an action over an approval limit, no progress possible.
• A request for a human means escalate immediately; frustration alone on a simple issue does not.
• Complexity by itself is not a trigger. Sentiment and self-reported confidence are unreliable signals.
• When ambiguity is costly, ask (e.g. for more identifiers); when cheap, assume and state the assumption.
• Gate high-stakes actions in code (tool preconditions or hooks), not only in the prompt.
• A handoff carries the reason, case facts, what was tried and a suggested next step.

Worked example: An insurance claims agent that escalates the wrong cases

A claims agent hands simple windscreen claims to staff but tries to approve a flood claim that falls outside the policy wording. The fix is explicit criteria in the system prompt (escalate anything outside written policy, any payout above the agent's limit, or any explicit request for a person), each with a short example of a case to escalate and one to resolve. A confidence score, a sentiment threshold or a separately trained classifier all skip the real problem: nobody told the agent where its authority ends.

Exam traps

• Escalating on negative sentiment or a low self-rated confidence score.
• Picking the most likely of several matching customers instead of asking for another identifier.
• Relying only on the system prompt to prevent refunds without verification.
• Investigating first when the customer has explicitly asked for a human.
• Escalating with a bare "customer needs help" and no case summary.

Official docs

• Effective context engineering for AI agents
• Hooks reference

Practise CCAR-F questions · Next lesson: Error propagation across agents