What actually prevents execution in agent systems?

About this article

Ran into this building an agent that could trigger API calls. We had validation, tool constraints, retries… everything looked “safe”. Still ended up executing the same action twice due to stale state + retry. Nothing actually prevented execution. It only shaped behavior. Curious what people use as a real execution gate: 1. something external to the agent 2. deterministic allow / deny 3. fail-closed if denied Any concrete patterns or systems that enforce this in practice? submitted by /u/docyb...