How do I revoke one AI agent without stopping the others?
Give each agent a separate credential or connection, revoke the one you want to stop, and test that the others still work.
Test both sides afterwards: the revoked agent should be refused, and an agent that still needs access should carry on working.
This is least privilege in practice. Grant each workflow only the access it needs, and make each grant removable on its own. The MCP authorization guidance describes protecting credentials and validating access tokens.
Try a two-agent access test
- Create two test identities or connections and grant each only the resource it needs.
- Make a harmless read with each one and record that both succeed.
- Revoke the first connection. Repeat the same read with it and confirm the service refuses.
- Repeat the read with the second connection and confirm it still succeeds.
- Look for other copies of the revoked agent's credentials and for active sessions. Revoking one connection does not prove that every unrelated credential is gone.
Use sample data for the test. Do not put real bearer tokens in logs or screenshots.
In ActAs, each agent has its own connection, and you revoke one without touching the rest. The guide to email access you can take back walks through it.