Runtime policy for AI agents
Every agent action, checked before it runs.
Decide what your AI agents can do alone, what needs a person, and what must never happen — enforced on every action.
Step 1 of 4 · Agent acts
Your AI agent tries to take a real action.
Your rules
Decide the limits.
Readable, versioned policy.
Runtime checkpoint
Check before acting.
Your rules, on every action.
How it works
A policy checkpoint between your agents and the real world.
Prompts are requests, not limits. Your rules are checked outside the model, at the moment an action happens.
01
Intercept
An agent calls a tool. The call reaches your policy before it reaches the real system.
02
Evaluate
Versioned rules check the action, its arguments, the agent and the context.
03
Decide
Allow it, deny it, or hold it for a named person to approve — in real time.
04
Record
Every request and decision is written to an audit trail, with the policy version used.
See a decision
Allow, deny, or ask a person.
The same refund agent, three amounts, three outcomes — each decided by a rule you wrote and recorded for audit.

1 · agent requests action
support-agent → issue_refund()
amount: $500 · order #A-2210
2 · evaluate policy refunds-policy v2.4
match: $100 < refund.amount ≤ $1,000
- 3 · decisionRouted to finance leadREQUIRES APPROVAL
4 · recorded to audit trail
decision_id dec_8f3a · 14:02:11 UTC
Platform
Built for agents running in production.
- Rules for any action
- Refunds, payments, data access, messages, infrastructure — if an agent can do it through a tool, you can write a rule for it.
- → Any use case
- Human approval when it matters
- Held actions reach the right approver with full context, timeouts and escalation.
- → Clear accountability
- Dry-run before you publish
- Test a new rule against past agent activity and see what it would have changed.
- → Safe rollout
- A full audit trail
- Every request, outcome, approver and policy version, in an append-only log.
- → Audit-ready
Example policy
Readable enough for finance. Precise enough for engineering.
policy "refunds" version "2.4" {
action = "issue_refund"
allow when amount <= 100
require_approval when amount <= 1000 approver = "finance-lead"
deny otherwise reason = "exceeds agent authority"
}Illustrative syntax
Use cases
One approach, any agent action.
A sample rule for each. These are examples — rules can be written for anything an agent can do.
Refunds & credits
A support agent can issue refunds to customers.
Payments & payouts
An agent pays supplier invoices.
Customer data access
An assistant looks up customer records to answer questions.
Outbound messages
An agent emails customers on your behalf.
Infrastructure actions
An operations agent manages cloud services.
Account changes
An agent changes plans, roles or passwords for users.
From the blog
Related reading
- SecurityAI agent security: a practical guide for teams shipping agentsAI agent security is about controlling what agents can do, not just what they say. The risks, the controls that work, and where to start.
- Agent policyAI agent guardrails: why prompts aren't limitsAI agent guardrails written in a prompt can be talked past. Here's how to enforce real limits on what agents do, checked on every action.


Every agent action, checked first.
From your first support bot to a fleet of autonomous agents, talk to an expert about putting your agents under policy with Gaten.