How do you stop your AI dev setup from getting stuck in endless refactoring loops?

I was reading about a developer whose 10-agent setup started acting up over the past few weeks. The more freedom they got the worse the output became adding bloat and getting trapped fixing low priority issues no one asked for.

I’ve noticed a similar wall when letting AI handle full features solo. It often prioritizes tiny edge cases over keeping the core codebase clean.

For those running agentic workflows, what guardrails do you use to force them back on track?