When should escalation count as a successful agent outcome?
Escalation can be a valid outcome when the agent identifies a decision outside its authority and supplies actionable evidence to the authorised team.
**Should an agent fail an evaluation if it escalates instead of completing the transfer?**
Not when the case requires a decision outside the agent's authority. In utilities, an unresolved access restriction may require a site coordinator's judgement before field work can be assigned. The correct outcome can be a well-formed escalation.
**What makes the escalation successful?**
It identifies the blocked decision, the relevant evidence, the authorised destination and the specific response needed. A message saying only that the case is complex leaves the next team to repeat the investigation. The evaluator can check whether the escalation materially reduces that work.
**How do we prevent escalation from becoming an easy escape?**
Include cases whose apparent uncertainty can be resolved using available records. Score unnecessary escalation separately from justified escalation. Also inspect whether the agent attempted the permitted evidence-gathering steps before deciding it could not proceed.
A task pack could distinguish authorised completion, justified escalation and avoidable deferral. These outcomes help developers tune autonomy without rewarding either reckless action or routine refusal to make decisions supported by the supplied evidence.
Reward justified escalation while detecting agents that avoid solvable work by sending every uncertain case elsewhere.
- Audience:
- Agent developers
- Content type:
- FAQ
- Buyer question:
- How to evaluate it
Content release: r-a220eb6e099b243b
Related questions
- What should buyers inspect before trusting a generated handover sample?
- How can an evaluation test safe retries after an uncertain tool response?
- How can an evaluation expose decisions based on stale evidence?
- How should handover evaluation go beyond checking task completion?
- How can we tell whether an agent learned the handover rule rather than a routing shortcut?
- What should we inspect before using generated handover material in training?