The hardest part of vibe coding isn't generating code—it's debugging phantom state

Generating an MVP with AI agents feels like magic until you hit day 3, when components start drifting out of sync.

When you're prompting full-stack features, agents are great at adding code, but notoriously bad at cleaning up after themselves. A few patterns I keep running into:

1. Redundant state loops: An agent creates three different local state hooks to track the exact same data across siblings, causing weird hydration mismatches and flickering UI.

2. Silent failure on edge cases: The agent writes an API call that works 95% of the time, but silently swallows errors in the catch block—leaving you staring at an empty UI with zero console logs.

3. Over-scaffolding: Asking the agent to "fix a layout glitch" often results in 40 lines of nested flex/grid overrides instead of tweaking a single container class.

When you're relying heavily on LLMs and agentic IDEs, what is your strategy for maintaining code hygiene?

Do you periodically ask the model to do a full refactor pass, or do you step in manually once an app grows past a few pages?