Agent Policy Simulator
See how policy and governance decide what an AI agent may do. Pick a scenario or build your own request, and watch the policy engine return Allow, Deny or Approval required, with the rule that decided it, the full evaluation trace and the audit record.
1. Pick a scenario
3. Policy decision
Evaluation trace
Every rule, in order. Deny overrides approve, approve overrides allow, and no match means deny.
| Rule | Effect | Result | Rule text |
|---|
Audit record
What the gateway would log for this decision. Metadata only, no payloads.
Edit the policy
Change thresholds, add agents or tools, or write your own rules, then apply.
Each rule has an id, an effect (allow, deny or approve) and when conditions that must all match. Conditions compare request fields such as tool, tool.kind, user.role, amount, instruction_source, target, delegation_depth, calls_per_minute or agent.external, using a value or operators >, >=, <, <=, !=, in and not_in. Try changing the payment approval limit from 10000 to 50000.
How policy and governance help AI agents
Without enforceable policy, most organizations keep agents read-only. Clear rules, enforced outside the agent, are what let agents take real actions safely.
- Scope first. An agent only sees the tools on its allow-list. A tool it cannot see is a tool it cannot be tricked into calling.
- Deny overrides. One matching deny rule, such as a blocked supplier or a recent bank-detail change, beats any allow.
- Humans approve high impact. Large payments and inventory postings pause for a named approver who is not the requester.
- Content is data, not instructions. A write action triggered by text in a document or partner message is refused as possible prompt injection.
- Default deny. If no rule says yes, the answer is no.
- Every decision is audited. User, agent, tool, decision, rules and trace ID are logged for every call.
Where these rules run for real
In production, rules like these are written in a policy engine such as Open Policy Agent (Rego) or Cedar, evaluated at the MCP or A2A gateway, and backed up by API management, integration flows and the authorization in each system of record.
FAQ
What is an AI agent policy?
It is a set of explicit rules that decides which agent may call which tool, for which user, on which data and up to which limit, and when a human must approve. It is enforced outside the agent, so a confused or manipulated agent cannot exceed it.
Why does a deny rule win over an allow rule?
Deny-overrides is the safest way to combine rules. Specific risk rules, such as a blocked supplier or a recent bank-detail change, must stop an action even when a general rule would allow it.
Is this simulator a real policy engine?
It is a teaching tool with a simple rule format that runs in your browser. Production systems use engines such as Open Policy Agent or Cedar at the gateway, with the same ideas: scope, deny-overrides, approvals, default deny and audit.