Replit's Agent Fixed a Bug, Broke Two, Fixed Those: My Loop
The agent iterates fast and sometimes regresses. The way out is smaller asks, not bigger ones.
I asked Replit's agent to fix one CSS bug and it did, then a button stopped working, then it fixed that and the layout shifted. The loop is real: the agent changes what it can see, not what you meant. I lost an hour to it.
The fix was shrinking the ask. Instead of 'fix the page,' I said 'only change the header padding, touch nothing else,' and it stayed in bounds. Big vague asks let it roam; tiny scoped asks keep it honest.
I also learned to read the diff before accepting. Replit shows what it changed, and skimming it caught two unrelated edits I could reject. The agent is fast enough that reviewing the small change is cheaper than debugging the surprise.
Used as a careful pair, it's great: I point, it types, I check. Used as 'make it good,' it's a regressions machine. The loop ends when the instruction gets specific.