# Production: Deployment, Observability & Cost

**Track:** Interoperability, Safety & Production — AI-Native Full-Stack Development — complete (30)
**Framework / surface:** web dev (CI/CD / operations)
**Level:** Advanced
**Prerequisites:** Evals, Tests, Traces & Regression, Durable Tasks, Queues & Background Agents

**In one line:** A production AI system is a governed service: deployable, observable, rate-limited, recoverable, and economically measurable.

## Theory, aesthetics & inspiration

A prototype proves a path can work once; production asks whether it keeps working under real users, failures, upgrades, and cost. The baseline is the discipline Google's Site Reliability Engineering (2016) codified — reproducible builds, protected secrets, staged environments, health signals, rollback paths, and error budgets that make reliability a number rather than a mood. AI adds operational variables ordinary web apps never had: model latency, token cost, provider rate limits, tool-call fan-out, background-agent duration, and quality drift when a provider silently updates a model. Measure them per successful user outcome, not per request — an agent that retries five times looks cheap per call and expensive per completed job. OpenTelemetry gives traces, metrics, and logs a vendor-neutral shape, and the capstone exercise is operational: deploy the full application, then break it on purpose — fail a tool, exceed a limit, reject an approval — and read the trace. The final literacy of this track is not building the system but understanding it while it runs.

**Builder question:** What signals would tell you that your deployed system is becoming slower, less reliable, less safe, or more expensive before users report it?