What I Learned Letting an AI Agent Refactor My Side Project

I've been putting off cleaning up an old side project for months. So last weekend I tried something different: I gave an AI agent the repo, one sentence of instructions, and stepped back to watch.

The setup

The project was a small Express API with no tests, three different date libraries, and a utils folder that had grown into a junk drawer. My instruction was simple: make it boring and consistent, and don't change behaviour.

What went well

• It found and removed two date libraries without me asking which one to keep.

• It wrote characterisation tests first, before touching any logic.

• Every change came with a short note explaining why.

What went wrong

It 'fixed' a rounding quirk that a client actually depended on. The tests it wrote encoded the new behaviour, so nothing failed. The lesson: an agent can only protect the behaviour you tell it about.

I'd do it again, but next time I'd write down the weird-on-purpose parts before handing over the keys.