Live AI Operationsaihabits.ai

Live AI Operations

The Stop Has to Live on the Credential

My halt took 5.9 seconds. Then 41 side effects happened anyway.

I stopped the main agent in 5.9 seconds.

Then 41 side effects happened anyway, the last one 58.1 seconds later. Thirty of them were permanent. Most did not come from the agent I stopped. A downstream agent worked through the queue it had been handed and produced 15 effects on its own.

Here is why. A child agent does not inherit its parent’s state. It inherits a credential and a budget. Once it has both, it has no reason to check back. Stopping the parent stops nothing the child can see.

The usual answer is to bound recursive delegation. Cap the depth, cap the children, cap the fan-out. That is worth doing and it does not solve this. A cap on the tree does not tell you which leaf is still spending. The leaf is not breaking a rule. It is holding a live token and doing what the token allows.

So put the stop somewhere the tree cannot route around it. One credential per agent. An expiry that is a date and not a policy. A revoke that hits that one agent and nothing else. Then the question of whether the stop worked becomes a list you can count: one credential, one row, one export.

That is a different test than the one most teams run. They stop the orchestrator and watch the orchestrator stop. Of course it stops. It was told to.

The test that matters is the ugly one. Pick a leaf, pull its credential mid-run, and see whether the effects stop. If they do not, you do not have a stop. You have a pause button on the thing you can see.

Five scheduled jobs fired in mine after I said stop. Nothing was broken. Nothing was unowned either. The jobs had a credential and a schedule, and neither of those listens to a person saying stop.

That is the whole point. Authority has to reach the thing doing the work, and the thing doing the work is rarely the thing you named.

Read this post on Substack, or subscribe for new issues.