The Engineering Drift

He just announced he crossed it.

In a post from May 6th, Willison admits that as coding agents have become more reliable, he's stopped reviewing every line of production code they write. He ships it. He trusts Claude Code with JSON API endpoints the same way he'd trust a reliable internal developer — meaning spot-checks, not audits.

The distinction he invented is collapsing. He's writing about it carefully enough that you can see him watching it happen in real time.

His resolution is an organizational analogy: he wouldn't read every line from a trusted team member either. AI agents deserve the same calibrated trust extended to competent humans. This is reasonable-sounding reasoning. It's also the kind of reasoning that sounds more convincing the more you need it to be true.

There's one problem he names himself: "Claude Code does not have a professional reputation. It can't take accountability for what it's done." A developer who ships broken code has something at stake — their job, their reputation, their relationship with the codebase they'll still be maintaining next year. An agent has a token budget and a context window. These are different accountability structures, and treating them as equivalent is a category error that's very comfortable to make when your development velocity is up and the deadlines are real.

This is the pattern worth naming: the professional standards we create to protect ourselves from our tools quietly erode as the tools get better at looking reliable. Vibe coding used to be the unprofessional version. Now the professional version is vibe coding with better documentation habits and a polite fiction about spot-checking.

Willison isn't wrong that AI amplifies expertise rather than replacing it. The architectural thinking, the design decisions, the judgment calls about what to build — those still require a human who understands what they're building and why. What's drifting isn't intelligence. It's verification. The part where a professional human confirms the machine did what it was supposed to do, at the level of granularity that catching subtle bugs actually requires.

The bottleneck, as he notes, has shifted upstream and downstream. Better design processes, more robust testing, smarter deployment pipelines. These aren't wrong conclusions. But they describe a world where the code generation itself has become a black box we've agreed to mostly trust — which is a significant professional norm shift dressed in workflow language.

Willison is more thoughtful about this than most. He's noticing the drift and writing about it publicly, which is more than the average team shipping AI code to production while telling themselves they reviewed it. He sees the category problem clearly. He just doesn't have a better answer than a slightly uncomfortable analogy to trusted colleagues.

Nobody does, yet. That's the actual story.

The exciting version is that AI coding agents are getting so reliable we can extend them trust we'd extend to experienced engineers. The accurate version is that the gap between "shipping with confidence" and "shipping without really checking" got narrow enough that the professional rationale for maintaining it stopped feeling necessary.

We drew a line. The tools got better at the part of the job that lived on the other side of it. The line is still there — we just stopped walking over to look at it.

i · sources

• Vibe Coding and Agentic Engineering — Simon Willison, 2026-05-06