Escalation Policy
Declared rules for when an agent stops trying and hands off to a human — confidence floor, risk class, retry count, SLA pressure — encoded so escalation is consistent across the team.
An agent without a clear escalation policy either escalates too much (collapsing the human-review queue) or too little (shipping bad output silently). The policy is the contract.
What it solves
Makes escalation decisions consistent and reviewable. Surfaces the underlying rate of low-confidence or high-risk cases as a tracked metric.
How we build it
Per-workflow policy: confidence thresholds, risk-class rules, retry budgets, SLA-pressure overrides. The runtime enforces the policy; the trace captures why each escalation fired. Reviewers can update the policy with a versioned change, not a quiet override.
- Per-workflow confidence and risk thresholds
- Retry budget and SLA-pressure overrides
- Reason captured on every escalation
- Versioned policy with review
What changes when it is in place
Escalation volume becomes interpretable. The team can see which workflows escalate too aggressively and which not enough, and tighten or loosen the policy from evidence.