An escalation is when an agent hands a task back to a human instead of completing it autonomously — because the task falls outside its scope, carries too much risk, or needs a judgment the agent should not make alone.
Escalation is a feature, not a fault: it is how human oversight is built into a deployment. But because each escalation is recorded in the ledger, the escalation rate becomes a measurable signal. It feeds the trust score, where a lower rate counts in the agent's favor, balancing autonomy against the cost of pulling a human in. A healthy agent escalates the right tasks without escalating everything.