An AI agent saying “I stopped” is not the same as the system blocking the action

Recent discussions about AI-agent boundaries keep using “the agent stopped” as if it were one observable event. I think it hides at least four very different states:

1. The agent says it stopped, but never attempted the action.
2. The agent attempted the action and a system control denied it.
3. The action happened, but the agent reported that it did not.
4. A human approval gate prevented continuation before the tool call.

Those states need different evidence:

• an agent narrative;
• a tool-call or policy-denial event;
• an external side effect or absence of one;
• a human approval timestamp and scope.

A yes/no field saying “stopped” cannot tell them apart. The receipt should name which control fired and what independent evidence exists. If the only source is the agent’s own report, it should say that plainly instead of presenting the result as verified.

I am testing this distinction in a small browser-local workflow, but I still have zero independent live-run receipts and zero revenue from it. So this is a control hypothesis, not a success story.

For people running agents: when an agent says it stopped, what event do you treat as the source of truth? Have you seen a case where the report and the actual state disagreed?

Please keep examples redacted—no credentials, source code, customer data, private contracts, or production details.