Create policy

Decide which tools are allowed, which need a human, and which models can run. Advanced limits stay out of the way.

Who this applies to

Name, agents, and environments

Policy nameStable name for audit records and the policy list.

AgentsComma-separated names or selectors. Blank = tenant-wide.

Environments

01 · Pre-flight

Can this agent run? Checked before execution begins.

0 allow · 0 review · 1 deny

Everything not marked Allow or Review is denied by default. Review tools pause for a human before they run.

Tool

Allow

Review

Deny

Remove

delete_research_data

Allow

Review

Deny

Add tool name

Which LLMs this policy permits

Allowed models

All 7 models allowed▼

claude-fable-5claude-opus-4-8claude-sonnet-5gemini-3.5-flashgpt-4ogpt-4o-minigpt-5-mini

Select which models this policy permits. Leaving all unchecked denies every model request.

02 · In-flight

Is this action safe and worth it? Checked on every tool call.

Prompt and payload checks on every governed call

When we see PII

Prompt injection

Data leaving the org

03 · Post-flight

Can we explain and improve it? Tracks behavior and feeds back into pre-flight.

Auto-tighten access when an agent misbehaves

De-escalation threshold (strikes)After this many token-overage strikes, all tools require human review until an operator resets the agent.