Pemicu Eskalasi yang Jelas

An autonomous Claude agent will eventually hit a wall it cannot or should not cross alone. The question isn't whether to escalate to a human, but when — and that decision must be defined by clear, deterministic triggers, not vibes.

In the Customer Support scenario, a refund agent that escalates too late frustrates users; one that escalates on every hiccup destroys its own value. This lesson nails down the four reliable escalation triggers the exam expects you to recognize:

• An explicit human request
• A policy gap (no rule covers the case)
• No progress after genuine attempts
• A threshold violation

If a customer asks to speak to a human, that is not a signal to interpret — it is an instruction to obey. Escalate immediately.

Do not try to talk the user out of it, re-route them through more self-service, or run sentiment analysis to decide if they 'really' meant it. A direct request for a person is the single most unambiguous trigger you will ever get. Treat it as a hard, deterministic branch in your control flow.

A policy gap occurs when the agent faces a case that no documented rule covers. Example: a refund request for a product category your refund policy never mentions, or an edge case the system prompt simply doesn't address.

The wrong move is to improvise a decision — the model will confidently invent a policy that may be wrong, costly, or non-compliant. The right move is to recognize the absence of guidance and hand off to a human who can make (and document) the call.

The agent should genuinely try before giving up. But once it has made real attempts — gathered identifiers, looked up orders, proposed solutions — and still cannot resolve the issue, lack of progress is itself a valid escalation trigger.

Note the distinction from a blind iteration cap. We escalate because the agent has demonstrably stalled on the task, not merely because a loop counter ticked over. The attempts must be real and observable, not a single failed call dressed up as 'no progress.'

Some actions exceed what an automated agent is permitted to do. A refund above a dollar limit, a bulk operation over a size cap, or any value crossing a defined boundary is a threshold violation — and a clean escalation trigger.

Thresholds are objective and auditable: the value either crosses the line or it does not. That makes them perfect candidates for deterministic enforcement rather than polite prompt instructions, as we'll see in the next scene.

For a threshold with financial, legal, or safety consequences, a prompt is not enough. Prompts are ~90% probabilistic; a determined or unlucky path can still slip past a soft instruction.

Use a PostToolUse / outgoing-call hook to block the action deterministically. The classic exam example: a hook that blocks any process_refund over $500 and forces escalation. Hooks give you 100% deterministic enforcement — the right tool when failure is expensive.

Just as important as the good triggers are the ones the exam flags as anti-patterns. Do not drive escalation from:

• Sentiment analysis — 'the user sounds angry' is noisy and gameable.
• Model self-rated confidence (e.g. 'I'm 6/10 sure') — LLM confidence scores are unreliable and uncalibrated.
• Untrained classifiers — an ad-hoc 'is this hard?' model you never validated.

These produce inconsistent, unauditable behavior. Stick to the four objective triggers: explicit request, policy gap, no progress, threshold violation.

An upset customer is not, by itself, an escalation trigger. The recommended pattern is: acknowledge the emotion → propose a concrete solution → escalate only if the request is reiterated.

This avoids two failures: coldly ignoring frustration, and hair-trigger escalation on every frustrated message. You de-escalate with a real fix first. Only if the customer repeats the demand for a human (an explicit request) do you hand off.

Escalation is not a dead end — it's a handoff. The human (or supervising system) needs context to act, so pass it in a structured form:

• Failure / trigger type (policy_gap, no_progress, threshold_violation, explicit_request)
• Attempted actions and queries
• Partial results already gathered
• Alternatives considered

A generic 'could not complete request' blocks recovery. Structured context lets the human resume instantly instead of starting from zero.

A subtle trigger-adjacent rule: when the agent cannot uniquely identify the subject — e.g. multiple customers match the given name — it must ask for more identifiers, never guess.

Guessing the wrong customer is a silent, high-cost error. This isn't an emotional or confidence judgment; it's an objective ambiguity in the data, handled by requesting disambiguation. Combined with identity-verification preconditions (block sensitive actions until get_customer returns a verified ID), it keeps the agent from acting on the wrong record.

A robust agent layers these triggers in priority order. An explicit request short-circuits everything. Threshold violations are enforced by hooks, not hope. Policy gaps and no-progress are model-recognized but escalated with structured context.

Crucially, escalation logic lives alongside the agentic loop, not inside its termination check. You still terminate the loop on stop_reason — never by parsing text — while these triggers decide when a human enters. Deterministic where stakes are high, model-driven where judgment is genuinely needed.

Key takeaways:

• Four good triggers: explicit human request (escalate immediately), policy gap, no progress after real attempts, threshold violation.
• Bad triggers: sentiment analysis, model self-rated confidence, untrained classifiers — noisy and unauditable.
• High-stakes thresholds (refund > $500) belong in deterministic hooks, not prompts.
• Emotion pattern: acknowledge → propose a fix → escalate only if the request is reiterated.
• Escalate with structured context (trigger type, attempts, partial results, alternatives) so a human can resume instantly.
• Ambiguous identity? Ask for more identifiers — never guess.

Objective triggers plus deterministic enforcement where it counts: that's architect-grade escalation design.