01 · Build the AI-native cloud¶

AI-native is a way of engineering, not a feature.

Putting a model behind an endpoint does not make a cloud service dependable. The engineering problem includes context selection, tool permissions, state, evaluation, deployment, and incident response. Design those parts together rather than treating safeguards as a post-demo task.

Start with an operational problem¶

Our scenario is incident triage: an operator wants to understand a degraded cloud service. The assistant may inspect a small, sanitised service snapshot and suggest a next step. It must not restart workloads, alter infrastructure, or invent missing telemetry.

The example deliberately separates a proposal from permission to act. It uses deterministic Python so you can test the harness without an API key. A production model would propose the same structured plan, which must be validated before dispatch.

from examples.cloud_triage import Incident, propose

incident = Incident(service="checkout", error_rate=0.12)
plan = propose(incident)
assert plan.action == "read_metrics"
print(plan)

0.12 means 12% of requests are failing. The initial policy uses 5% as a teaching threshold, not a universal production SLO. A real team derives thresholds from its service objectives and traffic profile.

Place AI in the Agentic SDLC¶

Use Copilot to explore the existing service and draft acceptance criteria before implementation. Keep the issue, reviewed specification, code change, tests, and operational evidence linked. The release decision belongs to the service owner.

Try with GitHub Copilot

Open examples/cloud_triage.py. Ask: "Explain the incident-triage flow without changing files. Identify the trust boundary between a proposal and tool execution. Draft acceptance criteria for a read-only first release, including missing telemetry and an unknown service."

Checkpoint¶

You should have a scope statement, a named release owner, and at least three observable acceptance criteria. One must say no infrastructure mutation. If the task is fully specified by stable rules, consider whether ordinary code is sufficient before introducing a model.