Where the work goes when an agent writes the code

After spending more time using agentic AI at work, I'm less interested in how quickly it can produce a patch than in what it takes to get that patch ready.

The typing is only one part. There's understanding the request, finding the relevant code, deciding what a sensible change looks like, checking the result and explaining it to somebody else. An agent can help with several of those steps. I still need to know which ones actually happened.

A confident summary is easy to read too quickly. I want to look at what changed and what was checked, then try the behaviour where appropriate. If there's a test, does it cover the problem? If the agent says a check couldn't run, that leaves me with a job to do.

I'm also noticing the difference between code I can read and code I understand well enough to maintain. A change can look plausible line by line while leaving me unsure why it took that approach. That's a reason to slow down, ask questions or simplify it.

This is useful work, but it uses attention differently from writing everything myself. I don't want to pretend that reviewing several streams of generated changes is free just because they were quick to generate.

For now, I prefer a task I can follow from the brief through to the result. I'll keep experimenting with the amount of room I give the agent. The measure I care about is whether we end up with a change we understand and are happy to own.