Approvals
When an agent is about to do something that changes the world (send an email, write to a system, spend money), it stops and asks. The card shows what will happen; hold the button to allow (long press on mobile) or choose Not now.
Under Settings > Approvals:
| Mode | Behavior |
|---|---|
| Always ask | Asks before any change. |
| Only what’s risky (default) | Reads and small edits just happen; asks before sending, posting, paying, deleting, deploying or running something. |
| Auto | An approval model allows what fits the request and asks about the rest. |
| Never ask | Runs everything. Use with care. |
In any mode you can allow or block specific tools and sites. A space can have its own rules, apart from the account’s.
Safety limits
Section titled “Safety limits”- Payment limit: in an agent’s settings you can set the most per payment and per month. Over the limit, or with no readable amount, the agent asks even if the tool is allowed.
- Unfamiliar tools: with an approval model set up, you can let a model suggest whether a tool with an unusual name only reads, makes a small edit or is risky. It is off by default, and when unsure the agent asks.
- Links: the agent will not open a link that carries a key, a very long query or text copied from private data, unless that exact link already appeared in the conversation.
Where they show up
Section titled “Where they show up”- as a card inside the conversation;
- under Approvals in the sidebar, listing everything you may decide across all conversations;
- in the inbox (the bell) and as a notification.
An agent with a pending approval does not continue until you answer.