Deploy an always-on agent — agent instructions

These are the instructions an AI agent follows for Deploy an always-on agent.

Goal

Build the documented FastAPI incident-desk service, give it a one-tool Dome identity, run it locally, then deploy it as a private Cloud Run service using an immutable container image and a Secret Manager reference. Invoke it and correlate the Cloud Run request with Dome Audit.

Rules

Show the active Dome context, Google account, project, and region before changing anything. Require a Dome sandbox and non-production Google Cloud project.
Create app.py, requirements.txt, Dockerfile, and .dockerignore exactly as published on the human tutorial page.
Never print or commit DOME_AGENT_API_KEY. Read it silently, pipe it into Secret Manager, then unset it.
Never put the key in source, a Docker build argument, image layer, ordinary Cloud Run environment variable, log, or chat.
Keep the service private. Do not use --allow-unauthenticated.
Keep /health local: it must not exchange a token, call Dome, or call the upstream tool.
Use an immutable image tag from the current Git commit; do not deploy latest.
Do not widen Cedar. The agent may call only demo-ops/it/get_incidents.
Show command output for every completed step, excluding secret values.
Do not delete resources until the user confirms.

Plan to present

Build and run the request-serving process locally
Provision sandbox-deploy-http, demo-ops, and incident-desk-api
Deploy and simulate the one-tool Cedar boundary
Create the API key and store it in Secret Manager
Build and push an immutable image
Deploy the private Cloud Run service
Invoke /health and /incidents
Correlate Cloud Run logs with Dome Audit
Show revision rollback and offer cleanup

Wait for confirmation before step 1.

Required evidence

Before reporting completion, show:

dome context current naming sandbox-deploy-http
one ALLOW for demo-ops/it/get_incidents and one DENY for another tool
local /health and /incidents responses
the immutable Artifact Registry image URI
the Cloud Run service account and secret reference, without secret content
a private Cloud Run service URL
one authenticated /incidents response
the corresponding tool.call in Dome Audit
the previous revision and exact rollback command

Interpretation

Explain that “always-on” means a request-serving contract, even when Cloud Run scales to zero. The application listens on $PORT, may handle concurrent requests, and lives across many requests. Cloud Run, Artifact Registry, and Secret Manager are implementation choices; the HTTP process, externalized credential, stable Dome identity, and one-tool rule transfer to equivalent platforms.

Do not claim the event-driven lab could use this image unchanged. A Lambda function needs a handler and runtime interface rather than a web server.