Interactive Sandbox
Follow the guided workflow below. No signup required.
The production Arbiter platform is currently under development. This sandbox demonstrates the governance workflow.
Step 1 of 5
Start with a preset to see how Arbiter governs real agent actions.
Finance Agent
Payment
Amount: $5,000
Expected
HOLD
Engineering Agent
Delete Database
Target: production
Expected
DENY
Support Agent
Send Email
Recipients: 3
Expected
ALLOW
Review the agent, action, and parameters — then evaluate.
When Arbiter returns HOLD, a reviewer approves or rejects before execution.
Evaluate an action that triggers HOLD to see the approval flow.
Every evaluation generates an immutable record.
Evaluate an action to generate an audit receipt.
One HTTP call between your agent and execution — like Stripe or Postman.
Updates live as you change agent, action, and parameters.
{
"agent": "finance-agent",
"action": "send_payment",
"amount": 5000
}Reflects Allow, Deny, or Hold after evaluation.
{
"decision": "...",
"reason": "...",
"policy": "...",
"receipt_id": "..."
}Drop Arbiter into your agent stack with a single evaluate call.
const decision = await arbiter.evaluate({
agent: "finance-agent",
action: "send_payment",
amount: 5000
});
if (decision.decision === "allow") {
executeAction();
} else if (decision.decision === "hold") {
await requestApproval(decision.receipt_id);
} else {
blockAction(decision.reason);
}Policies are declarative rules. Arbiter evaluates every action against them before execution.
If
Then
hold
If
Then
deny
If
Then
require_approval
Click each stage to understand how Arbiter sits between your agents and execution.
Your Agent
An autonomous agent initiates an action — payment, deployment, data access, or API call. It has identity but no inherent governance.
Join the Design Partner Program and help shape the future of AI governance.
This playground demonstrates the governance workflow. The production console is currently under development.