AI can now write code faster than anyone can read it, and platform engineering is suddenly relevant again. Two of its best-known practitioners say the old problems of capacity, testing and review are back, only bigger.
Ian Nowland, head of hardware at Baseten and an early AWS engineer, and Camille Fournier, VP of engineering at CoreWeave and former CTO of Rent the Runway, co-wrote a book on the subject. Both think coding agents have made a shared platform more valuable, not less.
Agents work better against a known platform
Fournier says the point at which a team outgrows owning its own tooling may now come sooner. People produce so much more, so fast, and it all still has to fit together. An AI builds quicker against a standard interface because there’s less room for it to make its own decisions and cause drift between teams.
Her worry is that vibe coding splinters in a million directions. Without constraints, you get something that works a million different ways across the company. Individual humans each driving their own AI need a common context, and she says that context is the platform, among other things. In her view, platform engineering supplies the constraints that keep agents consistent. She notes that after she and Nowland finished their book, AI took all the attention, and now people are asking what platform work was about.

CI is the new bottleneck
Nowland and Fournier agree that continuous integration is under strain as agents produce huge numbers of changes. CI compute, always expensive, is getting worse and at some companies is starting to rival production in cost.
Nowland says the earliest signs showed up at Anthropic, where tokens for developers on the latest models are close to free. The constraint was CI compute, test flakiness and the question of who switches the flags back on. A flaky test passes or fails depending on softer factors such as timing or the CPU it runs on, and it often breaks because of someone else’s change. Teams built flaky-test detectors, then turned the tests off and never turned them back on. He imagines AI working constantly to keep the number of disabled tests below a set level.
Centralized CI also pushes everything through a single pipe, which makes the queue a bottleneck.
Code review as a rubber stamp
Fournier says the question everyone is obsessed with is how to get rid of code review. Both she and Nowland grew up in an era when review was optional or reserved for sensitive code. Nowland estimates about 75% of the code he wrote in his first 11 years was never reviewed.
She traces the spread of mandatory review to Google and Microsoft culture, the move to Git and GitHub with its open-source habits, and Sarbanes-Oxley, which effectively requires review for code that touches financials. Regulators and auditors, she warns, have no sense of humor, so publicly traded companies need human review there.
Elsewhere, she argues, people already approve changes with a quick “looks good to me” because the volume is unmanageable. Her answer is to ask where human review is truly needed: regulatory and compliance cases and high-risk code, scored by another AI on factors such as security implications, production surface area and how long the code has gone untouched.
Testing is back to the heavy stuff
Unit tests, Fournier says, taught developers to write more testable code. When AI writes them, the tests may be well made but prove little, since AI tests what it thinks should be tested. She calls AIs satisficers: if the goal is to make the tests pass, they may do odd things to get there.
That makes full end-to-end integration testing the gold standard again. It’s expensive and compute-heavy, tests are slow and often flaky, and she expects teams to spend a lot of time chasing the failures.
The two-minute server was never free
Nowland points out that EC2’s pitch was a credit card, an API call and a machine two minutes later, rather than four months of procurement. It was partly a stretch, he says, because capacity behind it was finite. Someone still has to plan years ahead, ideally not the people at the edges.
Today that planning is under pressure. Fournier says a memory shortage about a year ago hit storage, servers and networking gear, lengthening procurement times and raising prices. On-demand capacity is scarce, and teams sometimes have to approach Google or CoreWeave for the first time to find it.
Her suggestion is a return to performance engineering. Teams may rewrite Postgres in Rust with agents in a month, and maybe it’s more memory efficient, she says, though she doesn’t know. Either way, people will be scrambling for resources like in the days when five teams shared a server. The cost shows up as money or as engineering time.
What teams can do now
The practical takeaway for platform engineering teams is to standardize the interfaces agents build against, decide which code truly needs human review, invest in end-to-end tests, and treat CI capacity as a budget item. Fournier and Nowland’s view is that the fundamentals of platform engineering haven’t changed. The volume of change flowing through them has.
Source: read.technically.dev