Forum Discussion
Difference between Copilot Agent settings and Microsoft Agent 365 Agent settings
Coralie, to make the verification suggested by Jamony concrete, I would use a small test matrix in a non-production scope: one user included in the access group, one excluded, and one agent assigned only to a restricted audience. Record separately whether each user can discover, install and invoke the agent, then test access to a permitted and a denied data source. Keep the agent, group membership and app policies fixed while changing one setting at a time through your normal change process. This helps identify which control is affecting the result without assuming the two screens share the same backend object.
In the governance model I published with PALO Framework (https://paloframework.org/PALO_AIGovernance.html), I connect each control to its scope, accountable owner, expected behaviour and observed evidence. For this case I would keep four rows: tenant access, individual agent assignment, data/tool permissions and spending eligibility. A successful access test in one row cannot validate the others.
Microsoft documents the centralized access control here: https://learn.microsoft.com/en-us/microsoft-365/admin/manage/agent-settings . If the two interfaces save different values, retain those observations for Microsoft support; PALO provides the evidence structure, not an answer about undocumented synchronization or Microsoft roadmap dates.