Exactly what happens, step by step

Sign up and create your agent in about a minute — no government ID required to start. It begins in unverified mode, which caps how much it can spend or commit to, regardless of what you later set its individual rules to. That ceiling lifts once you verify your identity, whenever you choose to.

Cautious asks before almost everything. Balanced handles routine things on its own and asks about anything unusual or costly. Autonomous handles more on its own within limits you set. You can fine-tune individual rules afterward — the style is just a sensible starting point, not a permanent choice.

Connecting Gmail or Google Calendar only grants your agent the ability to use that service — it does not by itself grant permission to send an email or move an event. Those are two separate checks. You can disconnect a tool at any time, instantly.

"Check my calendar tomorrow and move anything before 10." "Email Jennifer and tell her the quote is ready." Your agent reads the request, checks what it's actually allowed to do given your current rules and connections, and proposes a short plan before touching anything.

You see what it understood and the steps it intends to take. If a step needs a capability you haven't connected, or isn't something it can safely do, it says so honestly instead of guessing or making something up.

Reading your calendar, sending an email, spending money — each one passes through the same permission check, every single time, based on your current rules at that exact moment. This isn't something the AI model decides on its own; it's a separate, deterministic check that the model cannot override, skip, or grant itself.

If a step is above a threshold you set, or your rules say to always ask first for that kind of action, your agent stops and sends you an approval request. Nothing happens until you approve or deny it — and an approval is single-use, tied to that one action, never standing permission for next time.

Every check, allow, denial, and approval is recorded in a permanent activity log. You can pause your agent instantly, disable its ability to spend money with one tap, or revoke it entirely — which also cuts off any sub-agent it delegated work to.