Approval mode

How far the agent goes without checking in: Auto versus Manual.

The chip under the message box decides how far the agent goes without checking in.

Mode	Behaviour
Auto	The default. Does not ask. Fastest for exploration.
Manual	Asks before any create, update or delete. Use when touching live data.

You can change it mid-session. Switching to Auto after a few approvals is a common pattern: watch what the agent reaches for while it is unfamiliar, then let it run.

The approval gate

In Manual mode the agent stops and posts a card for each pending call. Nothing happens until you answer it.

One card per pending call. Each shows the exact tool name and the exact arguments.

Every card carries the same parts:

Part	What it shows
Title	The action in plain language, e.g. Create connect link
Tool badge	The literal tool being invoked, e.g. create_connect_link
Payload	The exact JSON arguments, such as the connectorId it will act on
VIEW RAW INPUT	Expands the full, untruncated payload
NOTE TO AGENT (OPTIONAL)	Free text: "On reject, this note is sent to the agent so it can re-plan…"

The three responses

• Accept: runs this one call, and only this one. The next call gates again.

• Always allow: stops asking for that tool for the remainder of the session. Useful once you have seen a read-only call a few times; think harder before using it on a tool that writes.

• Reject: refuses the call. Whatever you typed in the note goes back to the agent, which re-plans around it.

The note is what makes rejection useful. "Use the sandbox sheet, not the production one" produces a corrected plan. A bare Reject with no explanation usually produces the same call again.

Always allow is scoped to the session, not to the tool forever, but within a long session it can cover a lot of ground. If the agent is about to touch production data, leave the gate in place.

See the worked example for a gate in the middle of a real build.