Escalation: handing the task back

Lesson 5 of 5 in Human-in-the-Loop: Designing the Human Back In.

Gates operate on single actions. Escalation operates on the whole task: the agent concludes that a human should take over entirely, and hands it off. Every serious deployment needs this path, and the triggers are knowable in advance:

  • Repeated failure — the same subgoal has failed N times; more attempts burn money and goodwill.
  • A policy boundary — the task now requires an action the agent is scoped never to take (refund above the ceiling, legal commitment, HR matter).
  • Low confidence with high stakes — the agent can proceed but the cost of being wrong exceeds its mandate.
  • The counterparty asks — a user says “let me talk to a person”. Overriding that is how support agents make the news.
  • The task outgrew the plan — what looked like a password reset is actually a compromised account.

The design mistake is treating escalation as the failure branch — the sad path that gets a TODO comment. In production it is a success path: the system correctly recognized the limits of its mandate. An agent that resolves 70% of tickets and escalates 30% cleanly is deployable; an agent that “resolves” 95% by bluffing through the 25% it should have handed off is a liability with good metrics.

What separates an escalation the human thanks you for from one they dread is the handoff package. The receiving human was not in the loop; the agent was. So the agent’s last job is to transfer its context — not by dumping a 40-turn transcript, but by briefing:

The handoff package: what the human receives
ElementBad handoffGood handoff

State of play

“I couldn’t complete the task.”

What the task is, what has already been done (with effects — “refund of $120 already issued”), and what remains

What was tried

A raw transcript of every turn

The two approaches that failed and why they failed — so the human doesn’t retry them

The blocker

Nothing, or “an error occurred”

The specific trigger: which policy ceiling, which failing subgoal, which confidence gap

Suggested next step

Silence

The agent’s best guess, clearly labeled as a suggestion — the human decides, but doesn’t start cold

The customer’s state

Customer left hanging mid-conversation

Customer told a human is taking over, with the context already transferred so they never repeat themselves

Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.

Tool: Tool Permission Lab — Put the whole module to work: configure gates, scopes, and escalation paths for a live agent, then watch what your reviewers actually do under load.