Skip to main content

Concepts

Server and desktop access for AI agents: assignment, duration, approval and records

Granting AI agents access to servers and desktops: the assignment, duration, approval, revocation and record questions, and what Cerberus covers.

Browse all docs

Before an AI agent is given a job that touches a server or desktop, teams ask a simple question: on which resource, under whose authority, under what duration condition, and what is left when the work ends? The Cerberus operating model keeps task, resource, authority and duration defined; it takes required approvals from authorized users and ties decisions and results to reviewable work history. The assignment, approval, revocation and audit pages linked from here are written for user access; the identity and approval path an agent runs under is settled per setup.

Which questions come before giving an AI agent server access?

Five questions make the missing decision visible:

  1. Which resource will it work on? (One server or a group?)
  2. With what authority? (Is there more than the job needs?)
  3. For how long? (Was an end time or usage budget set?)
  4. With whose approval? (Which authorized person decides the work may start?)
  5. What can be reviewed afterwards? (Who, when, which decision?)

Which resource, with what authority and for how long does an agent work?

An agent is tied only to the resource assigned to it; other resources in the workspace stay out of scope. The time limit is optional: with timed access you set an end time or a usage budget, without it the assignment lasts until someone removes it. Assignment is covered in Identity, scope and assignment and duration in time-limited access.

When does an agent’s work wait for an authorized person’s approval?

When the active session policy requires approval, the work does not start until an authorized user approves it; not every resource requires approval. The decision rests with the authorized user, and the identity and approval path the agent runs under are settled before access is granted. Policy options are on the Session policies page.

How is an agent’s access revoked, and what happens to an open session?

When the work ends early or an agent’s task goes differently than expected, your team has two actions: removing the assignment and ending the open session. Removing the assignment does not close an open session by itself; the two are separate user decisions. The full sequence is on the Revocation and evidence page.

How is an agent’s work reviewed afterwards?

The records are event records: access request, decision, start, state change, revocation and close. For an agent’s work too, they show who requested the access, who approved it, and when the session started and how it closed. They do not show what was done inside the session; session or screen recording is in development. The Reviewing audit logs guide explains how to read them.

What does Cerberus not do for AI agents?

Limits:

  • It does not offer unrestricted autonomy. AI work stays behind assignment, session policy and record boundaries.
  • It does not control model behavior. What the agent produces or decides is outside Cerberus’s scope; Cerberus limits the resource, authority and duration.
  • It does not provide hosting. Cerberus manages access to your own servers, virtual machines and desktops; it does not rent machines.
  • It does not take over network responsibility. Network reachability and firewall rules stay with you.

The Cerberus Windows Agent is the connection component installed on a resource and has nothing to do with the AI agent. Resources connect through the Windows Agent or through agentless resource connection. On resources connected without the Windows Agent, in-machine readiness signals are more limited.

The live demo shows the same assignment, duration and revocation mechanics through a user session on sample data; the portal is read-only. You can talk to us about setting up an AI agent.