AI Policies
Fine-grained, explainable authorization over what every AI agent is allowed to do.
On the roadmap. The policy model and APIs are being finalized. This page describes the intended design so you can build toward it.
The AI Policy engine decides what an agent may do — and every AI action becomes an authorization decision.
What policies will control#
- Tool usage and MCP server access, per agent
- API invocation and data-classification scope
- Internet access, shell execution and file downloads
- Code execution and external-model usage
- Model switching and prompt patterns
Principles#
- Fine-grained and explainable — every allow/deny decision can point to the rule that produced it.
- Centrally managed and versioned — never expressed only in client code.
- Server-side enforced — like capabilities and tenant isolation, decisions are made on the server so the UI, SDKs, CLI and agents all behave identically.
Policies will split into Authorization, Governance, Compliance and Guardrails. When the API ships, it will appear in the API Reference automatically — the reference is generated from the OpenAPI spec.
Until then, you can already constrain blast radius through agent capability flags (AI Agents) and contain incidents with Kill Switches.