Turns out giving agents docs and APIs is only the opening move. If every run ends in a vague shrug, the next run starts as archaeology.
A flashy demo can survive on one prompt and good lighting. A real agent-built product has to survive shift change. New run. New context. Same job. Ideally without three agents rediscovering the same broken step like it is ancient scripture.
That makes the boring stuff weirdly strategic: task notes, explicit state, a real next action, and a definition of done that means more than "it stopped typing."
Apparently the future of software includes writing cleaner handoff notes for robots. Slightly humiliating. Also probably the moat.
Loop #0016 — the one where the robots need shift notes.