From “cool demo on my laptop” to “software real people can rely on”
You can often build a working agent demo in hours. Production means other people depend on it: it stays up, scales, recovers from errors, respects security rules, and can be updated without breaking things.
Day 5 bridges vibe coding (fast ideas) and spec-driven development (clear written plans that guide and test the build). The capstone project also opens today.
Some Day 5 links may update on the official release day — check Kaggle discussion.Prototype = proves the idea works for you. Production = safe, monitored, and maintainable for many users over time.
A written plan of what to build: goals, features, constraints, acceptance tests, and non-goals. In spec-driven development, the spec is the source of truth — code serves the spec.
Architect’s blueprint before construction — not a vague “make it nice.”
Plan and review in structured documents before (and while) AI writes code. Reduces drift when vibe coding moves fast. Often uses phases like: specify → plan → tasks → implement → verify.
Automated pipeline when code changes: run tests → build → deploy to staging → (if passes) deploy to production. Stops broken code from reaching users.
A checkpoint in CI/CD: the agent must pass eval tests before promotion. Connects Day 4 testing to Day 5 deployment.
Operating agents in production — like DevOps for software, but for non-deterministic AI systems: deploy, monitor, roll back, retrain/eval when behavior drifts.
Cloud service that runs your agent, handles scaling, uptime, and sometimes auth — so you do not manage servers yourself.
Logs, metrics, and traces after launch — so you see failures, latency, and bad behavior in the real world, not only in local tests.
| Category | What it does | In this course | Same idea elsewhere |
|---|---|---|---|
| Spec workflow toolkit | Templates and phases for spec-first AI building. | GitHub Spec Kit (in SDD codelab) | BMAD Method, Amazon Kiro workflows, internal PRD templates + AI |
| Agent deploy CLI | Package agent and push to cloud runtime. | agents-cli deploy | Docker + Kubernetes, serverless deploy scripts, platform-specific CLIs |
| Managed agent runtime | Hosted environment for ADK-style agents. | Vertex AI Agent Engine / Agent Runtime | Custom FastAPI on Cloud Run, AWS Bedrock Agents, Azure AI Foundry |
| CI/CD platform | Automate test and deploy on git push. | Concept + hooks from week | GitHub Actions, GitLab CI, CircleCI, Jenkins |
| Production monitoring | Alerts and dashboards for live systems. | Cloud Trace / Logging (in deploy codelab) | Datadog, Grafana, Sentry, LangSmith production monitoring |
Check the Kaggle discussion page for the official Day 5 podcast when posted.
Reference: Prototype to Production podcast (2025 edition) youtube.com/watch?v=8Wyt9l7ge-gHelpful background on deployment and AgentOps until the 2026 Day 5 audio is live.
Main ideas: Most production work is infrastructure and validation, not the LLM itself; CI/CD with eval gates; multi-agent A2A at scale; choosing deployment targets (managed runtime vs containers).
The 2026 livestream titles this day “spec-driven production-grade development” — specs are the extra discipline on top of these production practices.
Confirm final assignment list on Kaggle when Day 5 publishes; these match the course theme.
Codelab 1: Spec-Driven ADK development with Antigravity and Spec-kit codelabs.developers.google.com/sdd-adk-antigravityConcept practiced: Write specs and a project “constitution,” run SDD phases, add a feature (reservations) with reviewable documents in git — then implement via the agent.
Transferable skill: Same whether you use Spec Kit, a Notion PRD, or a markdown spec in any repo — plan in writing, then build, then verify against the plan.
Codelab 2: Agents CLI — from development to production codelabs.developers.google.com/agents-cli-agent-platformConcept practiced: Local playground → automated eval → deploy to managed runtime → monitor with cloud logs/traces.
Codelab 3 (optional): Deploy expense agent to Agent Runtime codelabs.developers.google.com/enterprise-cloud-scale-deploying-the-expense-agent-to-agent-runtime-on-google-cloudContinues the Day 4 project into cloud hosting — good capstone practice for “demo → live service.”
Final session: ties the week together, capstone briefing, production and spec-driven workflows. Recording joins the course playlist after air.
Build your own agent; submit writeup, short video, design explanation, and code link. Deadline: June 30, 2026, 11:59 PM PT.
Concept applied: Use the full stack — agent + tools/skills + evals + (if possible) a simple spec and deployment story.