What can an always-on agent do without asking for approval?

OpenAI’s dots act on their own except on anything that could touch your accounts or share information, which routes through an automatic check called auto-review before it runs; Google’s Gemini agent instead grants access up front and checks the result afterward, with no per-action sign-off at all.

Dots’ auto-review weighs a pending action against your own instructions, your Custom Rules, and OpenAI’s safety requirements, then decides what proceeds on its own, what needs your approval, and what you have to do yourself; a few tasks, like changing a password, always stay with you. Gemini skips that live check: you grant it role-based access once, a policy firewall called Agent Gateway blocks whatever that access shouldn’t reach, and every action still lands in an audit trail you read after the fact, with a per-project spend cap as the one thing that pauses it until you click resume.

So one vendor decides live, action by action; the other decides upfront, through what it was given access to. Either way, a clean audit trail only proves an action happened: nothing in a log like that checks whether the result was actually right.