Giving an agent a job it can finish

Using agentic AI at work has made me pay more attention to how I describe a task. Something that feels obvious in my head can turn out to be several decisions I've never written down.

“Sort this out” leaves a lot of room. Which behaviour is wrong? What should happen instead? Is there an existing pattern to follow? I need answers to those questions whether I'm doing the work myself or asking an agent to take a pass at it.

I'm finding a bounded change easier to work with. Give the agent enough context to inspect the relevant code, explain the outcome, and specify how it can check the result. Then look at the diff. That last part still takes time, especially if the agent has found a more elaborate solution than the problem needed.

There's a temptation to expand the job once it gets going. While it's in there, it could tidy this, rename that, update something nearby. I have the same temptation when I'm writing the code. The review gets harder either way.

I want to be able to explain the finished change without referring to the conversation that produced it. If I can't, I need to do more work before it's ready. Having an agent write the code doesn't remove that responsibility.

I'm still learning where this is useful. The interesting part is how much the quality of the brief affects the quality of the work. It makes vague instructions rather difficult to ignore.