I see a lot of excitement around agent loops. AI agents can now continue working on a task, review their own work, use other agents for feedback, and keep iterating until they reach a result.
I use AI agents all day, every day. They have completely changed the way I write software. My experience with long-running loops is a bit different from what I often see online.
I think they work well for short, well-defined tasks. Things like classifying emails, simple refactors, dependency updates, or other work where it is easy to measure whether the result is correct.
For larger development tasks, I get better results by staying involved.
The issue is not that today’s models are not capable enough, but that software development is rarely a straight line. You learn while building, requirements change, and better ideas appear. Sometimes you discover that the original approach was wrong.
Because of that, I don’t like waiting until the end of a long loop before looking at the result. I usually find it more effective to adjust the direction every few minutes. Those small corrections prevent larger corrections later.
I know many loop-based workflows already include human checkpoints. I think that makes sense. My preference is simply to make those checkpoints much more frequent than many examples suggest.
There is also a practical reason. Long-running loops consume a lot of tokens, and adding more review agents increases that cost. In many cases I would rather spend a few minutes steering the process myself.