Building reliable software with AI agents

You asked an agent to build something, and the software it created ran the first time you tried it.

Two years ago this process would have required a developer, a sprint and a budget, and now it needs an afternoon.

The trouble usually arrives with the first change you ask for, when the thing that worked on day one suddenly stops working. Working once and continuing to work are different properties, and in our experience only the second one is worth paying for.

The build loop: plan in phases, write failing tests, independent reviews, smoke test in browser, update memory — The build loop: five steps, repeated once per phase.

Enter your email to read