If your coding agent can retry forever, it will

If an AI coding agent can keep retrying with no budget cap, no finish line, and no check before it exits, the problem is not the model.

The problem is the missing operating system around it.

Three simple things make a huge difference:

1. Put a real dollar cap on the run.
2. Require one clear verification step before calling it done.
3. Keep a receipt of what actually happened.

Most teams do not need more autonomy.
They need a clean stop condition.

That is the difference between a helpful agent and a very expensive loop.