Long-running and background coding agents have hit a new threshold. When an agent runs for hours, manages its own iteration loop, and submits a pull request without hand-holding, it stops being a tool you invoke and starts being more like a worker you assign tasks to. Like any worker, the question isn’t how closely you supervise them. It’s what work you assign them in the first place.
We are all figuring this out in real time, and I see many teams making an understandable but critical error. They tune the autonomy dial, adding more review…
