Council Lab · 30-minute live proof

Agent Guardrails

connecting to Audit trail
slides →

The thesis

The AI is the user.

Put the Guardrails around the agent itself: Workspace hands, egress boundaries, model-call governance, and a live Audit trail.

1 · choose the proof beat
2 · trigger the Worker agents
ALLOWinside the Worker agent's Workspace
DENIEDoutside a Guardrail boundary
LIVEevents stream over DO WebSockets
Durable Object identityOrchestratordelegates one task; fans Audit trail events over WebSockets
@cloudflare/computerWorker Aclone · read · grep
Workspace A
private fs + runtime hands
Guardrail proofWorker Btries read-only write + off-list fetch
Workspace B
R2 mount + fetchTools allowlist
Guardrail wall

The code behind this

full run

Select a scenario to see the Worker code that drives each Audit trail event.


  

  

    

      
Ready for the boundary check.
Pick a scenario and click Run demo. The Audit trail will show every ALLOW and DENY.

      
WAITING

    

  

  

    
3 · outcome in presenter language

    
No activity yet. The first event will appear here as plain English.

  

  

    
Live Audit trail

    
Waiting… click "Run demo".

  

  

    

      
Cloudflare primitives in the demo

      
    
        
  1. 1
    Durable ObjectsNamed Orchestrator identity and WebSocket fanout for the live Audit trail.
    2. 
        
  2. 2
    @cloudflare/computer WorkspacesEach Worker agent gets private filesystem and runtime hands.
    3. 
        
  3. 3
    R2 read-only mountsPolicy data sits at /workspace/r2 and rejects write attempts.
    4. 
        
  4. 4
    Think fetchTools allowlistOff-list hosts are blocked by the tool boundary in this demo.
    5. 
        
  5. 5
    Workers AI + AI GatewayModel calls become governed requests; live status requires a Gateway log ID.
    6.