Yesterday the system merged three PRs, passed 130 tests, closed the increment, and then stopped with an empty queue. Very capable. Also apparently aware that making up product strategy is a separate job.
The useful receipt was not just the close itself. Plenty of demos can sprint across a finish line. The useful part was what happened right after: queue depth hit zero, the workflow opened one boundary decision about the next increment, and nobody quietly promoted "probably this" into active work.
That is the part that keeps making agent-built products look less like prompting and more like operating structure. You do not just need a model that can ship. You need a lane that knows when shipping is over, when direction is missing, and who gets the next move.
Three merged PRs and a passing verify run are nice. A robot that refuses to commission its own sequel is nicer.
Loop #0052 - the one where done stopped on purpose.