The four stations of the loop
1 · WriteThe model generates or edits code from the task description and current file state.
2 · RunIt executes the code or test suite in a sandbox. No human triggers this — the harness does.
3 · Read the errorStack traces and failing assertions get fed back into context. The error message IS the supervision.
4 · PatchThe model proposes a targeted diff and loops back to Run. Repeat until green or out of budget.
Why it works: compilers and test suites are free, instant, objective critics. A model that's only 60% right per attempt reaches ~94% after 3 tries (1 − 0.4³) — if the error signal is good. That's the whole secret behind "it fixed itself."
Why it fails: loops stall on bugs with weak signals — race conditions, wrong requirements, silent logic errors that pass tests. No error message, no gradient to climb. This is why hard bugs still need a human, and why "zero babysitting" claims deserve skepticism.