Which triggers should a municipal agent treat as appropriate escalation conditions?

Municipal agent escalation should be driven by clear, auditable conditions rather than vibes. The appropriate triggers are an explicit human request, a documented policy gap the agent cannot lawfully fill, and inability to make progress on the resident's goal despite available tools and playbooks. Those three conditions keep 311 and permitting desks honest: they escalate when the resident asks, when the knowledge base has a hole, or when the loop is stuck—not when a topic merely feels hard.

This triad works because each trigger maps to a concrete operational signal. Explicit request is resident consent for handoff. Policy gaps prevent the agent from inventing rules on tax, code enforcement, or eligibility questions. Inability to progress covers tool failures, contradictory records, or exhausted recovery steps without pretending every tough topic needs a human by default.

Treating any vaguely complex topic as an escalation condition fails because complexity is subjective and would flood clerks with cases the agent could still resolve under policy. Escalating only after three failed tool calls and never on policy gaps fails because a missing ordinance answer is already a policy-gap escalate even before tool retries, and arbitrary retry counts are not a substitute for playbook holes. Escalating only when sentiment analysis scores the resident as angry fails because tone models misread accents, terse civic speech, and disability-related communication; anger is not a reliable proxy for escalation need.

Exam caveat: appropriate triggers are request, policy gap, and stuck progress—not sentiment, not vague complexity, and not a single magic tool-failure count. Operational check: before escalating a municipal case, record which of the three triggers fired and attach evidence (quote of the human ask, missing policy cite, or failed recovery attempts).