Trust the tests, not the vibes.

I didn’t fuss much about verification when I first started building with AI. The agent was good enough that I felt I could trust it, including the assumptions it made along the way.

I’d ask it to fix something, then that fix would break something else, and we’d keep going. Assumptions stacked on top of assumptions until I had AI slop and an audit trail so large I couldn’t possibly follow it. Eventually, I couldn’t confidently explain all the changes we’d made.

That’s why verification is the most important part of building with agents for me. I think a bad plan with good verification can take you further than a good plan with bad verification.

When I say verification, I mean things like unit tests, integration tests and browser automation. Those checks need to cover the behaviour that already worked as well as the thing we just changed. Otherwise, it’s easy to keep fixing the latest problem while losing track of what broke along the way.

People say tests matter less now because agents can write good code. I don’t think that’s true. Codebases grow over time, and no one can guarantee that an agent is holding every edge case in its head at once. Deterministic tests built up over time get very close to that, which is why they matter.

Tell the agent what good looks like and give it a way to check its work, and it can take you most of the way from there.