Approvals and human in the loop

An agent with a Gmail connector could send an email to the wrong person. Human in the loop means a person approves certain actions before they happen.

For each connector on an agent, Tool Management sets which tool calls (actions) pause for approval:

• Always allow: everything runs without asking (the default).
• Ask each time: every call, even reading, needs approval.
• Ask for writes/deletes: reading runs freely; anything that creates, changes, sends or deletes needs approval. Gumloop recommends this as a good default.
• Custom: choose per tool, including Never allow.

Advanced app rules can add conditions, such as "ask only when the email goes outside our company".

When an approval is needed, the agent pauses and saves its progress. You see an approval card showing the tool, what the agent intends, and the key values (recipient, subject, amount). Click Approve or Reject. You can also respond from the notification bell, from the Tasks page (filter "Approval Required"), or with buttons in Slack.

Separately, the Ask Question ability lets the agent stop and ask you to choose between options before continuing, which is useful when a request is ambiguous.

Read approval cards properly. Clicking Approve without reading turns the safety net into decoration. Start strict, and relax rules for specific tools once the agent has proved itself.