Do not start a loop you cannot check.
The agent optimizes whatever the check measures, so a weak check gives confident wrong work.
Loops
The habits that make the difference. Tick the ones you already keep and see what is left.
0 / 4 you keep
Write a stop condition as evidence and add a hard cap.
Stop on tests passing, not "until it looks good", and cap turns, time or cost. Same failure twice means stop and explain.
Take one task per iteration and keep state on disk.
A plan file, a progress file and git history let each fresh session pick up where the last one stopped.
Separate the maker from the checker, and retest the harness with each new model.
Every harness part encodes an assumption about what the model cannot do, and that assumption goes stale.