The agents area is governed by capability, not by job title.
| Action | Who can |
|---|---|
| See the list of agents | anyone in the workspace |
| Chat with a published agent | member or above |
| Create, configure, publish, schedule | administrator or above |
| Connect outside systems and approve tools | whoever administers the workspace |
| Authorize an action | whoever has permission for that action |
The rule that matters most#
Authorizing what the agent proposed requires permission for the action itself — being able to chat with it is not enough.
If you can't reprocess a pipeline from the menu, you also can't authorize the agent to do it. The agent is not a shortcut to permissions you don't have.
This applies down the chain: when a coordinator asks a specialist, the permission is checked again, and it is still yours.
A high-risk action requires two people#
Whoever proposed it doesn't authorize their own proposal. It's the same idea as the person who pays not approving their own payment.
Scheduling acts with the permission of whoever scheduled it#
An automatic run has no authority of its own: it uses the permission of the person who created the schedule, checked on every trigger. If that person is deactivated, the schedule stops and says so.
Nothing crosses between customers#
Each workspace sees only its own agents, tables, documents, and runs. This doesn't depend on configuration — it's how the system behaves.