Platform · How it works
Between the agent and the real world.
Gaten sits in the path of every tool call. Nothing reaches a real system until your policy has answered.
Your agent
Any framework or model
Gaten
Allow · approve · deny
Your systems
Payments, CRM, cloud, email
One request, end to end
Follow a $500 refund through the checkpoint.
- 1
Intercept
The agent calls a tool. Before the call reaches your refund system, CRM or cloud account, it goes to the policy checkpoint.
Tool call agent support-agent action issue_refund args { order: "A-2210", amount: 500 } user customer #88412 - 2
Evaluate
The published version of the matching policy checks the action, its arguments, the agent, the user and context like amount, region or customer tier.
Policy match policy refunds @ v2.4 rule 1 allow when amount <= 100 ✗ rule 2 require_approval when <= 1000 ✓
- 3
Decide
One of three answers comes back — allow, deny, or require approval — with a structured reason the agent can act on.
Decision { "decision": "require_approval", "approver": "finance-lead", "reason": "above auto-approve limit" } - 4
Record
The request, decision, approver and policy version are written to an append-only audit trail.
Audit entry 09:12:04 support-agent issue_refund $500 HELD → APPROVED by finance-lead policy refunds@2.4
Connect your agents
Two ways in. Same decision.
Wrap tool calls in code, or route tool traffic through a gateway. Either way, the agent gets the same structured answer.
SDK
Call the checkpoint from your tool wrapper, then act on the answer.
const decision = await policy.check({
agent: "support-agent",
action: "issue_refund",
args: { order: "A-2210", amount: 500 },
});
if (decision.allowed) await refunds.issue(order, amount);API gateway
Point the agent's tool traffic at the gateway. Calls are checked before they're forwarded.
# agent config TOOLS_BASE_URL=https://policy-gateway.internal # allowed calls are forwarded unchanged; # held calls wait; denied calls return a reason
Illustrative code
Fails closed
When in doubt, the action doesn't run.
- Checkpoint unreachable
- The action is denied, not waved through.
- Approver doesn't answer
- It escalates, then is safely denied.
- No policy matches
- Denied by default — unless you choose otherwise.
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.
- SecurityPrompt injection in AI agents: how it works and how to contain itPrompt injection turns AI agents against their owners. How direct and indirect attacks work, and why the best defence sits at the action layer.
- EngineeringDeploy AI agents to production: a readiness checklistBefore you deploy AI agents to production, check identity, limits, approvals, testing, monitoring and rollback. A practical readiness checklist.


See it on your own agents.
An expert can walk through a real tool call from your stack, end to end.