Step 6 · Wrap-up: production and next steps (10 min)

Goal: Learn how the system goes to production, and what you can do next.

You built the full loop: instrumented app → tools → ADK multi-agent SRE agent → Orchestrator with a policy → chat UI with A2UI surfaces.

Production on Cloud Run (facilitator demo)

./bootstrap.sh   # pick/create a project, link billing, store GEMINI_API_KEY in .env
./deploy.sh      # APIs, least-privilege service accounts, 4 Cloud Run services (~6 min)
./cleanup.sh     # delete everything again

The important part is the IAM split:

Service account	Can	Cannot
sre-chaos-monkey-sa (target app)	Write traces and logs	Read telemetry
sre-agent-sa (SRE agent and Orchestrator)	Read traces, logs and metrics. Write its own trace spans. Use Firestore. Read the GEMINI_API_KEY secret.	Change telemetry or infrastructure

The target app makes incidents, but it cannot read telemetry. The SRE agent investigates incidents, but it cannot change telemetry or infrastructure. It only writes its own trace spans. Thus, the SRE agent can also diagnose its own requests.

This is least privilege at the IAM layer. The policy from step 4 adds least privilege at the agent layer.

⚠️ The demo deploys the services with --allow-unauthenticated. This makes the chat easy to open. After the session, delete the deployment (./cleanup.sh). Alternatively, put the services behind IAP before you share the URL.

Use the Antigravity skill

.agents/skills/sre_incident_solver/ contains the same SRE agent as an Antigravity Agent Skill (SKILL.md + code). scripts/sync_skill.py generates the skill from sre_agent/. Thus, the skill always agrees with the code that you built.

Antigravity finds the skills in .agents/skills/ when you open this repository:

1. Run uv run simulate_incident.py --engine-only. It makes an incident.
2. Open the repository in the Antigravity app, or run agy in it.
3. Ask: "Use the sre_incident_solver skill to diagnose the latest incident."

The agent reads SKILL.md and runs the diagnosis (it asks before it runs a command). This is a third way to use the same engine: in the chat (A2A), in a terminal (simulate_incident.py), and as a skill of a coding agent.

The same folder has the workshop skills: sre-agent-setup, sre-workshop-coach and sre-agent-deploy.

Next steps

EXERCISES.md contains follow-up exercises in three levels:

Level	Exercises
1: Tools and reports	Make the metrics show the truth, log-pattern clustering tool, SLO and error-budget section, severity in the post-mortem
2: Workflow and agents	Mitigation Planner agent, multi-trace incident correlation, human-in-the-loop remediation tool, notification to Slack or a webhook
3: Architecture and production	RAG over runbooks, live on real GCP, evaluation harness
Capstone	The fully autonomous loop: alert → diagnosis → post-mortem