Escalation: when a human has to decide
Last updated: · Leutrim Miftaraj
Escalation (step-up) is the mechanism that reconciles autonomy and oversight: below the threshold the agent acts alone; above it the person who granted the mandate is asked — with context, a timeout (600 seconds) and deny as the default.
The mechanics
Verify returns escalate with an escalation ID; the person approves or rejects; the verifier verifies again with the escalation_token. Approvals can be consumed exactly once — a token authorises one action, never a series.
Why before the reservation
An escalated request reserves no limit (OMP v0.5.2): only the confirmed re-verification books against the monthly limit. Pending step-ups therefore never block legitimate parallel actions.
AI Act link
Art. 14 requires effective human oversight — the documented escalation history is its proof: who was asked when, what was decided, what happened next.
Key takeaway
Step-up is not friction; it is the condition under which people are willing to give agents authority at all.
Try it yourself
Issue your first agent mandate and check your first decision with its receipt — free, in a few minutes.