Handoff / Escalation
Clear escalation criteria are what prevent an agent from getting stuck or acting beyond its authority.
In one sentence
Handoff protocol: when the agent transfers the task to a human or to another agent. Defining clear escalation criteria is what prevents an agent from getting stuck or acting beyond its authority.
Before
Agent reaches the limit of its authority
What it does
Escalation criterion triggers
After
Human or another agent takes over
Example
Escalation rules for the letters agent: report with contradictory information, child who left the program, any mention of a risk situation, and verification failure twice in a row.
In these cases it does not try to resolve it: it stops, writes the reason, and calls the coordination team.
The common mistake
Leaving the criterion implicit, such as "escalate if you are not sure". The agent has no certainty calibration. It under-escalates, especially in the cases where it should escalate more, because those are the cases where it does not know what it does not know.
In practice
- Write the triggers as a closed list, in code. “If you are not sure” is too vague to work as a trigger.
- Whoever escalates provides context: what they tried, where they stopped, what they already know.
- Measure the escalation rate. Too low is as suspicious as too high.
How to make it tangible
A written list of escalation triggers, implemented in code and outside the prompt.
Translated from Portuguese with AI assistance.