One alert.
Three realities.
No silent gaps.

1. An authorized advisory arrivesModel Armor screens the text inline. A prompt injection or personal identifier is blocked before any model reads it.
2. Gemini reads it, and writes the planOnly fields it can quote word for word survive. Embeddings order the playbooks, Memory Bank recalls these facilities, and Gemini synthesizes tasks that must cite a public-health source.
3. A planner agent delegatesIt chooses which specialist runs next and why, or holds while replies are still plausible. The registry refuses anything outside the registered table.
4. Durable wakes are registeredAcknowledgement deadlines live in Firestore, not in a browser tab. Cloud Scheduler fires them on wall-clock time.
5. Facilities answer, or they do notReplies arrive by console, API or Pub/Sub. Gemma triages them and Gemini checks any photo against the claim. Silence is escalated by the wake itself.
6. A human decides what only a human mayScarce-resource allocation, advisory issuance and rescission. Then recovery is verified, memory is written, and a briefing is drafted.

Five specialized agents, five identities

• incident-plannerDecides which specialist the fleet delegates to next, in what order, and when to wait instead. Executes nothing itself.
• facility-fleetActivates the registered fleet against an authorized advisory.
• policy-gatewayRejects actions outside registered authority and delivers only pre-authorized playbooks.
• resource-coordinatorSurfaces scarce-resource conflicts and escalates silence. Allocates nothing.
• recovery-verifierVerifies facility recovery, and only when an authorized rescission arrives.

Each runs on its own Vertex AI Agent Engine runtime with its own Agent Identity, behind one governed Agent Gateway, catalogued in Agent Registry. A plan may only name registered role and command pairs; compliance steps are enforced by the registry, and a planner outage falls back to the registered sequence.

Catalogued, and bounded to one region

• Cross-department discoveryAll five capabilities are published to Google Cloud Agent Registry, so another department finds and reuses them without redeploying this fleet. A sixth facility class is a registry entry, not a rewrite.
• Weeks, not sessionsAcknowledgement deadlines, a 42-day recovery wake and Memory Bank facts outlive every session, browser and container instance.
• Data residencyRuntimes, incident state and wake state stay in one project and one region (us-central1). The public service accepts fictional exercise data only.
• Least authority by constructionOne identity per agent, one governed gateway, inline screening on every supplied text, and three decisions no agent can reach.

Running on Google Cloud

• Gemini 3.5 FlashQuote-verified extraction, adaptive playbooks, photo evidence, after-action briefing
• Gemini Embedding 001Semantic playbook routing
• Gemma 4Facility reply triage through the Gemini API
• Google ADKAll five agents, planner and specialists
• Agent EngineFive scale-to-zero runtimes and Memory Bank
• Agent Identity + GatewayOne identity per agent, one governed access path
• Agent RegistryCapabilities published for cross-department discovery
• Model ArmorInline screening of every supplied text and model output
• Cloud RunThe orchestration service you are using now
• FirestoreVersioned incidents and transactional wake state
• Cloud SchedulerOIDC-signed background wake scans every minute
• Pub/SubAsynchronous facility event feed
• Cloud TraceEnd-to-end request correlation
• Secret ManagerAPI-key pepper and model credentials