Some decisions should not be made by an agent. When an agent reaches one, it can stop and ask a person, an escalation. The run pauses where it is, waits for an answer, and then carries on with that answer in hand.

The Order Fraud & Refund Decision agent is the clearest example: it decides small, unambiguous cases itself and escalates anything high-value or borderline, according to the thresholds you set when you installed it.

What happens when an agent escalates

  1. The agent's run pauses. Its invocation shows the termination reason Waiting for a human response.
  2. A notification goes to the install's escalation target, the escalation user if one is set, otherwise the escalation group, with a Respond action.
  3. If the install has a chat channel configured, the question is posted there as well.
  4. When someone answers, the run resumes from where it stopped.

An escalation waits indefinitely. Nothing times it out, and the job stays parked until it is answered, so an unattended escalation queue is a stalled automation, not a self-clearing one.

Answering

The Respond to agent screen shows:

  • Question: what the agent is asking, in its own words.
  • Agent reasoning so far: how it got here, so you are not deciding blind.
  • Your answer: either a set of options the agent offered, or a free-text box. Options include Something else — describe it below, and any answer can carry extra context through Add detail.
  • Recent agent activity: the transcript up to the pause.

Submit response sends your answer and releases the run.

Detail is worth adding. The agent continues with your answer as input, so "reject, the delivery address is a known reshipper" produces better follow-on reasoning than "reject" on its own.

After answering

The screen becomes read-only and records Answer given, Answered by and any Detail you added, with a View agent run link through to the full invocation.

The notification changes its action to View response, so anyone else who was notified can see the decision was made and by whom rather than answering it twice.

Who gets asked

Escalation targets are set per install, on the install wizard:

  • Escalation user: one person. If a user and a group are both set, only the user is notified in the admin panel.
  • Escalation group: everyone in the group, when no user is set.
  • Escalation chat channel: always notified in addition to the admin notification, never instead of it.

For an agent that escalates regularly, a group is usually more reliable than an individual: one person on leave should not stall a queue. Use the chat channel to make sure escalations are noticed promptly, since nobody watches the notification bell continuously.

If nobody can answer

There is no way to cancel a waiting escalation from the escalation screen. The practical options are to answer it (any answer releases the run) or, if the whole run is no longer wanted, to leave it parked and re-run the job later from the hook batch.