Policies
Define how agents are allowed to act.
Set rules for tool use, risky action categories, egress, model usage, and approval thresholds.
Live
Policy Editor
Author the live policy for governed agent actions.
This is the policy that actually runs. Edit tool controls, risky categories, egress allowlists, and approval thresholds.
Open Policy EditorSandbox
Policy Lab
Test governed compile behavior before you ship it.
Preview how a candidate policy behaves on real prompts, before promoting it to live.
Open Policy LabReview impacted actions in Action Control
After saving a policy, see how it affects pending and recent requests.
Run a Proof Check
Confirm the policy still passes evaluations and audit integrity checks.
What goes in a policy
Tool controls
Per-tool and per-category rules: allow, ask, deny, or redact.
Risk categories
Default behavior for read / write / delete / send / network / credentials / code / admin.
Approval thresholds
Risk levels above which an action must pause for human review.
Egress allowlist
Domains an agent may reach when making network requests.
Model allowlist
Providers and models the agent is allowed to call.
Argument patterns
Regex constraints on tool arguments for tighter scoping.