AI Agent uses two complementary controls for connected services: connection access limits what an agent may use, while approvals let a person review proposed external changes before execution.
How the controls work together
Limit available access
Assign only the connected accounts and tools needed for the agent’s role.
Let the agent prepare work
The agent can use its allowed reads and propose an allowed external change within its instructions.
Review the exact proposal
When approval is required, inspect the provider, account, action, and displayed arguments before deciding.
Execute or stop
Approval allows that proposal to execute; rejection stops it. Neither decision silently broadens the agent’s access.
Review access over time
Recheck connection access when an agent’s role changes, a teammate leaves, an account is replaced, or a workflow begins making different kinds of changes.