AI agents can build a feature in an afternoon. The question is who checked it. Every feature we build goes through the same fixed loop: spec, build, review, a short verdict, and one human who decides what goes live.
An AI agent can build a feature in an afternoon. It writes the code, it runs, it looks fine. And then the question nobody likes: who checked it?
What we saw
If AI writes the code, who's responsible when it breaks? Who looked at it before it touched your invoices?
There are two easy answers and both are wrong. Let the agent ship its own work and you get speed with no safety net. Check every line by hand and you lose the reason you used AI in the first place. We wanted the speed and a person in charge, so we gave every feature the same route.
What we do
We don't use one big AI that does everything. We use several AI agents, each with one job, run in a fixed order. Each agent only gets what it needs for its own step, so the spec writer doesn't build and the builder doesn't review its own work. Every feature of any size goes through the same loop:
- A spec writer turns the request into a spec: what's in, what's out, and numbered acceptance criteria. "Done" means something before anything gets built.
- A builder builds it on its own branch, away from the code that's live.
- Our own checks run first (tests, type checks, a full build). If one is red, the reviewers don't even start.
- Review agents go over it: code quality, security, and when there's a screen, the UX in a real browser.
- Whatever they find goes back to the builder. At most two fix rounds, then it moves on, with the open points written down.
- An approver reads everything and writes a short verdict: ready or not, how risky, and up to three things worth a look.
- Then it's my turn. Only I merge. Going live is a separate step with its own yes.
Where money is involved (prices, invoices, VAT), I also approve the acceptance criteria before the builder starts. The merge is the only door to production, and I'm the only one with the key.
Agents checking agents isn't a guarantee, and we don't pretend it is. It catches a lot of the boring and the obvious. The last call is still made by someone who knows the client and carries the responsibility.
A rule of thumb
Ask whoever builds your software with AI one thing: who decides what goes live, and what did they look at first? If the answer is "the AI, mostly", you're getting speed on credit. If the answer is a person with a short list of what was checked, you're fine.
What about you?
Using AI to build, or thinking about it, and not sure who's holding the pen at the end? Let us come and look. The first look is free, no strings.

