The Engineering of Intent, Chapter 32: Vibe Coding in Platform and Infrastructure

This is Part 32 of a series walking through my book The Engineering of Intent. In the previous chapter, we covered data and ML. Chapter 32 turns to platform and infrastructure — the tooling other engineers depend on.

A Different Character

Platform work is relatively slow-moving; its users are internal; its correctness is measured by whether other engineers can do their jobs without friction. AI-native practices apply here too, with specific adaptations — and with some of the highest leverage I see anywhere, because every improvement compounds across the whole engineering org.

Three Subdomains

1. Infrastructure as Code. Agents produce correct Terraform syntax; they’re erratic at correct abstractions. Every resource tagged with owner, environment, purpose. Automated scans enforce. Hunt and consolidate ad-hoc modules that almost-but-not-quite duplicate existing ones.
2. CI Pipeline Evolution. Agents can identify redundant steps, parallelize stages, cache artifacts. What they cannot do reliably: prioritize between flakiness fixes and feature development. That’s a cultural decision. The practice that works: allocate one engineer-week per quarter exclusively to CI improvement. Publish the before/after numbers.
3. Internal Tools. The ideal Vibe Coding pilot domain. Ship the first version in two days. If it doesn’t ship in two days, the scope is wrong. Ugly but functional; users love it; iterate for two weeks; at the point of acceptable-to-quite-nice, stop and build a different tool.

The Deployment Tool Rebuild

A platform team I advised had a 2019 deployment tool. It worked, but was slow, error-prone in edge cases, and opaque in progress reporting. They’d been planning a rewrite for a year.

“AI-native rewrite: six weeks, two platform engineers plus agents. Week one: interviewed ten engineers about pain points, produced a prioritized feature list, decided on a new state-machine architecture. Weeks two and three: core state machine, five most common paths. Week four: less common paths, remediation tools. Week five: integration testing, shadow parallel runs, user test. Week six: rollout and deprecation of the old tool.”

Deploys from engineer-initiated to green-in-production dropped from 40 minutes average to 12 minutes. Engineer satisfaction on “deployment experience” went from a median of 5/10 to 8/10.

Next up — Chapter 33: Building Your Personal Vibe Coding Operating System. The final chapter of Part IX zooms out from domain deep-dives to the personal layer — the setup, habits, and artifacts that turn practices from this book into your own durable system.