“Done” is rarely a property of the work itself. It is a judgment produced by the system around the work. Humans rely on a patchwork of signals like tests, specifications, precedent, approval, deadlines, and finding that point of diminishing returns. Completion comes from outside the work itself.

An AI model can almost always produce another answer. It does not become tired of the work. This makes @steipete’s reminder compelling: you shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents. Instead of manual prompting, systems run cycles themselves, checking results and deciding next steps. However, the loop is only as good as the verifier at each step. Even before we started talking about loop engineering, everything already runs as a loop, just with a very expensive tool call – human hand prompting and serving as the verifier.

Take the standard coding-agent loop: keep working until the tests pass. But tests are only a proxy. In Spec Bench, frontier agents routinely passed visible tests while failing held-out tests. One agent produced a 2,900-line “compiler” that simply memorized inputs. The loop converged on the verifier, not intent. Loop engineering is not merely retrying. It is making each cycle reduce the distance between the current state and a desired state. A loop is not yet a direction.