Platform · Policies
Rules your finance team can read. And your engineers can trust.
A policy says what an agent may do alone, what needs a person, and what must never happen — in plain, versioned rules that both sides sign off.
Anatomy of a policy
Read it line by line.
11policy "refunds" version "2.4" {22 action = "issue_refund"343 allow when amount <= 10054 require_approval when amount <= 10006 approver = "finance-lead"75 deny when customer.flagged8 deny otherwise reason = "exceeds agent authority"9}
- 1
Named and versioned
Every saved change is a new version. Decisions always point to the version that made them.
- 2
Scoped to an action
One policy per tool or action, so owners can find the rules that affect them.
- 3
Safe by default
Small, reversible actions run automatically. Nobody waits for a $20 refund.
- 4
A person when it matters
The middle band is held for a named approver, with context.
- 5
Hard limits
Some things never happen, whatever the prompt says. The agent gets a clear reason.
What a rule can check
More than the amount.
Rules can use anything about the request and its context.
| The action | issue_refund, send_email, delete_resource |
|---|---|
| Its arguments | amount, currency, recipient, environment |
| The agent | which agent, which version, which team owns it |
| The user | who the agent is acting for, and whether they're verified |
| Context | customer tier, region, time of day, daily totals so far |
Change control
From draft to live, with a record at every step.
- 01Draft
- 02Review
- 03Dry-run
- 04Publish
- 05Roll back
Dry-run before you publish.
Lower the auto-approve limit from $200 to $100? Replay last month's refunds through the new version and see exactly which actions would change, before anyone's queue fills up.
| Action | Was | Would be |
|---|---|---|
| #A-1932 · $180 | allow | require approval |
| #A-2047 · $95 | allow | allow |
| #A-2210 · $640 | require_approval | require approval |
| #A-2391 · $1,250 | require_approval | deny |
From the blog
Related reading
- Agent policyAI agent policy: how to write rules your agents actually followAn AI agent policy sets what agents may do alone, with approval, or never. How to write one, who owns it, and how to enforce it on every action.
- EngineeringHow to test AI agent guardrails before productionTest AI agent guardrails by replaying real agent activity through a new rule before it goes live, so limit changes never surprise anyone.
- SecurityAI agent permissions: deciding what an agent is allowed to doAI agent permissions are more than API scopes. How to decide what each agent may do alone, with approval, or never — and enforce it per action.


Turn your limits into rules.
Bring a process your agents touch, and an expert will help you write its first policy.