There are two ways an agent stops asking permission for something: it earns the right by getting the same thing right repeatedly, or you grant it directly.

This article covers:

1. Earning a permission
2. Granting one directly
3. Why it still asks

Earning a permission

An agent builds a record by being useful without being corrected. Approve the same kind of action — say, updating an activity — ten times in a row without asking for changes, and that action becomes eligible.

Eligible does not mean granted. When the agent reaches ten, it does not quietly start acting on its own. It raises a permission request, and a person still has to approve that before anything changes. So there are two gates: the agent earns its way to the request, and you decide whether to grant it.

Note: the Permissions tab says the permission "unlocks automatically" at ten. That wording is misleading — what happens automatically is the request, not the grant.

Two more details:

• Only a clean Approve counts. Using Request changes means the proposal was not right, so the count starts again.
• It is per action, not global. Ten clean approvals of "update activity" makes updating activities eligible. Creating contacts is earned separately.

Ten is the default; an administrator can change it.

Granting one directly

If you already trust the agent with something, skip the wait:

1. Open AI Workspace → Agents and click the agent.
2. Click Open agent page.
3. Go to Permissions.
4. Under Entity access, click + Allow something now.
5. Choose the record type (for example Activity) and the action (Create, Modify or Delete).
6. Click Allow.

The agent can do that one thing unattended from then on. Everything else still comes to you as a proposal.

Tip: grant the narrowest thing that solves the problem. "Modify activity" is a much smaller promise than "Delete company", and you can add another later.

Why it still asks

• The action is not the one it earned. Create and modify are separate, and so is each record type.
• You have been requesting changes. Each one restarts the count.
• It is a capability, not entity access. Sending email and similar actions sit in the Capabilities section and are granted separately — see What an agent can do on its own.
• Wrong agent. Permissions belong to one agent. A new agent starts read-only however well another has performed.

Related

• What an agent can do on its own
• How to approve, edit or decline what an agent proposes