AI Didn't Replace Developers — It Replaced Certainty

Here's the uncomfortable truth about AI in the build loop: it didn't remove the hard problems — it removed the comforting ones.

Boilerplate is cheap now. Scaffolding is cheap. Even a convincing first draft of architecture is cheap. What got expensive is trust.

You can watch a model invent a solution that compiles, passes a happy-path test, and still misunderstands your domain by one quiet assumption. That is the new craft: not producing more code, but interrogating what the code claims to know.

This essay maps the full loop — requirements, design, implementation, tests, review, ship — and asks where human judgment must interrupt generation.

Requirements are where most of the damage is scheduled. If the prompt is a vibe, the PR will be a vibe with types. Write the outcome first: who it is for, what must never happen, what "done" looks like in production rather than in a screenshot. Then let the agent fill volume.

We argue for a split you can defend under pressure: AI owns volume; humans own intent, constraints, and consequence.

"It works on my machine" now includes "it worked in the agent's story." The agent's story is a closed world with a tidy narrative and a missing webhook. Verification is the scarce skill of 2026 — not because tests became fashionable again, but because generation made false confidence cheap. A passing unit test on a generated payment helper is not a threat model. It is a compliment.

Design interruptions are cheap compared to production interruptions. Stop generation at the boundary of identity. Stop it at money. Stop it at the migration that rewrites history. Let it run wild on the button, the empty state copy, the test name. That is not Luddism. That is taste with a pager.

Strong teams build rituals around that scarcity. Acceptance criteria before prompts. Threat models before payments. A named reviewer on identity and money paths who is allowed to say no without a meeting. Refusal is a professional muscle. If you cannot explain the change to the person who will get the 2 a.m. call, you have not finished — you have only generated.

If AI feels magical, you're looking at the output. If it feels like a responsibility upgrade, you're looking at the work.