Access, control and trust

Common questions about Agent access, permissions, approvals and trust.

What systems and data can the Agent access?

Only capabilities available to the current run.

That depends on your access, Project or Channel scope, configured Connections and the permissions of the external account behind those Connections.

Where do passwords and API keys go?

Use the supported Connection and credential flow.

Do not paste secrets into a conversation, Skill or Project instructions.

Can the Agent change our ERP, CRM or other systems?

Yes, when an authorised Connection exposes the required action and the current run is allowed to use it.

Consequential actions can require approval.

How do approvals work?

The Agent can pause for your judgement, and configured write actions can require explicit approval.

A typed reply can answer a pending write-access approval. It does not bypass the approval.

What happens when the Agent gets something wrong?

Correct the current task, then identify the cause if it is recurring: the Skill, source data, Memory, Knowledge, Connection or Workflow.

Leave feedback when useful so builders can inspect the bad outcome with the execution evidence.

How do we verify important work?

Check the actual post-condition in the system that owns it—the ERP record, sent email, created file or other business result.

See Permissions and approvals and Working well with the Agent.