What actually stops an AI coding agent from looping?

[Keesan12]

One pattern I keep seeing in long-running coding-agent setups: teams add more visibility, more prompts, more retries, or more cost dashboards, but still do not have a clean answer to one runtime question.

Should the next attempt run at all?

That decision usually needs at least five pieces of state:

• remaining budget
• retry / iteration count
• whether the verifier outcome improved
• the last failure class
• explicit halt reason if the run should stop

Identity, MCP, CI, and billing all matter, but they sit adjacent to that decision rather than inside it. By the time CI reports a failure or billing reports the spend, the bad loop already ran.

The runtime seam that has felt most load-bearing for us is attempt -> verify -> record, with admission checked again before each next attempt. Without that, a system can have good tools and still burn money in circles.

Curious where people are putting that decision today:

• inside the agent runtime
• at the tool boundary
• in the orchestrator
• in CI / workflow logic
• somewhere else entirely

If you are already running coding agents unattended, what has actually worked in practice?