During an incident, outbound agent tools for the resident assistant must be disabled quickly. What control fits?

During an incident, outbound agent tools for the resident assistant must be disabled quickly. The fit control is a sentinel kill-switch feature flag in Azure App Configuration that agents check before calling outbound tools. When the flag flips to ‘tools disabled,’ municipal agents skip tool invocation, fail closed on external actions, and continue safer degraded modes (for example answering from cached policy text) while responders investigate. Feature flags propagate faster than image rollbacks and are reversible when the incident ends.

Waiting for the next quarterly release train to remove tool code from every image fails the urgency requirement: quarterly cadence cannot contain an active tool-misuse or downstream outage. Unplugging the city’s entire internet uplink as the only mitigation fails proportionality—it punishes unrelated civic systems and is a crude network hammer rather than an application control. Deleting the App Configuration store so no service can start until the outage ends fails as incident response: it expands the blast radius, destroys the control plane you need for flags, and turns a tool pause into a citywide AI outage.

Exam caveat: a kill switch only works if agents evaluate it on a short refresh interval and treat missing or fail-open configuration carefully—prefer fail closed for dangerous tools. The flag should gate tool execution, not merely logging. Pair the switch with monitoring so ops confirm tool call rates drop after the flip.

Operational check: in staging, enable the outbound-tools kill switch, confirm resident-assistant tool spans stop and errors are controlled, measure time-to-effect after the App Configuration change, then disable the switch and verify tools resume without redeploying agent images.