Permit
Decide exactly what every agent can do.
Each agent holds a list of capabilities, and each entry is one of three things: allowed unattended, allowed after a named human agrees, or never. The default for everything is never. You grant upward from nothing rather than restricting downward from everything.
Deny is the default, not the exception
A new agent can do nothing at all. Every capability it ends up with is one somebody added deliberately, which is why a finished agent's permission list is short and readable.
Some things are never grantable
Deleting a customer, issuing a refund, changing another agent's permissions. These are refused at the platform rather than left as a setting an administrator could switch on at four in the afternoon.
Capability and authority are separate
Holding the tool that changes a price does not mean being allowed to offer any price. The policy engine checks the second thing after the permission system has allowed the first.
In this area
6 pages under Permissions.
Each one is its own surface in the product, and each one is here because it does something the others do not.
Grants and scopes
Allow, approve or never — per capability, per agent.
ReadAuthority
Holding a tool is not the same as being allowed to use it freely.
ReadBoundaries
What an agent can never do, however it is configured.
ReadPermission reviews
Access granted for a reason that expired is the common failure.
ReadSecrets
Agents never hold credentials. Ever.
ReadNetwork policy
Where an agent may reach, and nowhere else.
ReadThe rest of the platform
This is one part of one loop.
Give AI a job.
Build an agent, choose what it can access, define what it may do, and deploy it to work safely across your business.
No code required · Full SDKs and API · Stop any agent instantly · [email protected]