Approving what agents write
Agents can read your connected systems freely, but before one creates data, changes data or sends a message in another system, it stops and asks you. The request appears as a card in the chat where the agent is working.
Answer an approval
The card says what the agent wants to do, which connector it will use, and what kind of change it is: Creates data, Modifies data or Sends message. Open the details to see the values it will send.
Choose one of:
| Choice | What happens |
|---|---|
| Allow once | This change goes ahead. The agent asks again next time. |
| Always allow → Allow connector | This change goes ahead, and the agent stops asking for that connector. |
| Always allow → Allow all connectors | This change goes ahead, and the agent stops asking for any connector. |
| Decline | Nothing is written and the agent stops. |
When the agent asks to run several changes together, one approval covers them: You approved this action chain. Later writes in this turn will not ask again.
If nobody answers
The card shows Waiting for your approval with the time left. If nobody answers within 10 minutes, the agent does not make the change and the card says Timed out. Ask again to retry.
If you stop the turn before answering, the card says Cancelled and nothing is written.
Where approvals appear
Answer an approval in the chat where the agent is working. A view that cannot answer it, such as a read-only transcript, says Open the chat this turn ran in.
Approvals for a whole workstream's outputs, such as a report before it is published, work differently: see Reviews and commits.
FAQs
Does reading data need approval? No. Approvals are for creating, changing and sending.