Coordination has a price

When there are multiple agents, a job appears that did not exist before: deciding who does what.

That is the orchestrator’s role, and it is business logic. A complete report goes to the cheap model. A contradictory report goes to the expensive one. A child outside the program goes to no agent at all, but straight to a person. Written as an explicit rule, this often cuts the average cost in half without anyone noticing a difference in the outcome.

The second new job is the handoff. Every agent will reach a case it does not know how to solve, and what happens at that moment determines whether the system is reliable or dangerous.

Escalation is the part people get wrong most often. The classic mistake is leaving the criterion implicit, such as "escalate if you are not sure". The agent has no certainty calibration: it under-escalates, precisely in the cases where it should escalate more, because those are the cases where it does not know that it does not know.

An escalation trigger is a closed list, written in code. And whoever escalates hands over context: what it tried, where it got stuck, what it has already found. Escalation without context gives the whole problem back to the human, and the automation ends up saving nothing.