How do you stop an always-on agent from going off track?

You tell it, mid-run: OpenAI’s dots let you step back into work already underway and redirect it, and its own docs say you can return with new information or priorities “without restarting the work.” You can also ask what it’s working on before deciding whether to intervene.

That’s correction after the fact, not a gate up front. Custom rules work the same way, as “ongoing boundaries” rather than a one-time check at the start: checked before each new action, not applied back onto one already taken. Noticing that a long-running agent has started doing something new is what tells you there’s something to redirect.

Google’s Gemini agent has no equivalent dot-style check-in. It writes every action to an audit trail you can monitor in real time to catch anomalous behavior, but that’s a person reading logs, not the agent reviewing itself; the only automatic lever is the per-project spend cap, which triggers on cost, not behavior.

Neither vendor’s pause undoes what already happened. OpenAI’s docs say stopping a task “doesn’t undo actions it has already completed,” so catching an agent going off track still leaves the earlier part of the run to clean up by hand.