Rule
When an agent hits ambiguity, missing information, or repeated failure, it stops the line and escalates. Guessing forward is how small confusions become large incidents.
The supervisor sits on the loop, not in it: testable intent goes in at the top, an escalation signal comes out when a worker pulls the cord, and bounded retries live in between. Escalation triggers are defined before work starts — ambiguous requirements, a failed attempt budget, an unexpected permission ask. Accountability is the one thing an agent cannot hold; the tree exists so every stop lands on an owner who can.
What AgentOps Enforces
- Give every worker one supervisor and one obvious escalation path.
- Define stop conditions up front: ambiguity, attempt-budget exhaustion, surprises outside the brief.
- On repeated failure, hand the failure context to a fresh agent — don't loop the stuck, saturated one.
- Keep humans at the top of the tree for judgment calls, not inside every retry.
Failure Signal
- An agent resolves ambiguity by picking the interpretation that lets it keep going.
- Work bounces sideways between peers until nobody knows who owns the decision.
- The same failure retries all night because stopping was never an option.
- Humans discover the incident from the damage, not from an escalation.
Done Looks Like
When a worker gets stuck or confused, the line stops fast, the escalation lands on one owner, and a fresh agent — not the saturated one — picks up the retry.