Giving an agent a job it can finish

Using agentic AI at work has made me notice how many decisions I leave out of a task description because they seem obvious to me.

"Sort this out" could mean fixing a bug, changing the design or replacing the implementation. Before handing it over, I need to say which behaviour is wrong and what should happen instead. An existing example in the codebase helps explain what I have in mind.

Then I need a way to check it. Describing the expected result makes the diff easier to review, especially if the agent has written more code than I expected.

I find it tempting to expand the job while it's running. It could tidy this, rename that, update something nearby. I do the same when I'm writing the code myself, and it makes the review harder either way.

Before the change is ready, I want to be able to explain it to someone who hasn't read the conversation with the agent. If I still need that conversation to understand why the code looks this way, there's context missing from the change.