Watch the control model in action. Introduce an incident, detonate the kill switch, and see how a human decision becomes a coordinated shutdown.
BROWSER SIMULATION · NO LIVE CONNECTIONS
Illustrative sequence. Animation timing is not a product benchmark.
01 / CHOOSE YOUR SCENARIO
KILLSWITCH / HUMAN CONTROL PLANEDEMO ENVIRONMENTRUNNING
CONNECTED SYSTEM
agent-workflow-01
Execution path Human authority
◎ HUMAN OPERATOR FINAL SAY
⏻
KillSwitch control planeOutside the agent’s authority
AGENT EXECUTION BOUNDARY
✳
Primary agentRUNNING
Research agentRUNNING
Execution agentRUNNING
TOOLSCONNECTED
API KEYSVALID
NETWORKOPEN
CONTAINED / HUMAN HOLD
EXECUTION NODES3 running
ACCESS PATHS3 connected
RESTART AUTHORITYHuman only
03 / THE SHUTDOWN SEQUENCEAWAITING HUMAN COMMAND
01
Accept overrideA human makes the call
✓
02
Stop executionRoot & child workloads
✓
03
Block accessTools & network paths
✓
04
Revoke credentialsNo authority to act
✓
05
Hold & verifyEvidence. Human review.
✓
ILLUSTRATIVE EVENT LOGSESSION READY
HUMAN AUTHORITY REMAINS AVAILABLESIMULATION ONLY / NO LIVE AGENTS
04 / BEHIND THE SWITCH
One decision. A coordinated response.
The product direction, made visible. Each enforcement path needs a real integration.
CONTROL
A separate authority.
The stop command originates outside the agent’s own permissions. Human control remains available after agent execution is halted.
CONTAIN + REVOKE
Cut the ability to act.
Stop connected workloads, block their tools and network paths, and revoke credentials at the issuer. Verify each path independently.
RECOVER
A deliberate restart.
Keep incident evidence, review the cause, and require human approval before restoring a workload’s execution or access.
↳
A real kill switch only covers integrated systems under your authority. Unconnected agents, independent copies, and completed actions need separate response procedures.