My coding agent can write the code. Why am I still its project manager?

The thing that bugs me about coding agents isn't even the mistakes anymore. It's how much managing I still have to do. I can explain a task, watch an agent write a bunch of code, and still end up being the person who notices it skipped something. Or that it stopped halfway through. Or that "done" apparently meant "I've finished talking for now." I've seen people describe the same thing. One developer had an actual step-by-step plan, but their agent kept breaking the work into smaller pieces and stopping until they typed "continue." No new instructions needed. No decision to make. Just... continue doing what I already asked. And yeah, I expect to review important code. That's not really my complaint. It's having to constantly remember the original task for the agent and decide whether it has actually finished. I've been messing around with a small open-source experiment because of this. The idea is to keep the completion checks separate from the agent doing the work, so its own "done" message isn't enough. Someone else tried it on a small task. Codex stopped before creating the required file. The check caught it, sent the task back, and Codex finished on the second attempt. Nobody had to type "continue." Obviously one small test doesn't prove much. The checks can miss requirements too, and right now the whole thing only handles a narrow set of tasks. Still, it got me thinking that maybe knowing when you're done is a different problem from being good at writing code. I'm curious how this works for people using agents on bigger projects. Have you actually had an agent finish a substantial task while you were away, without coming back to type "continue" or point out something obvious it missed? I'd love to hear what worked, or a specific time it didn't.