
Every agent request decided by Open Policy Agent
Argent ships Open Policy Agent inside its gateway. Every model call, tool call, agent-to-agent call and agent action is checked against Rego rules before it leaves, and anything the engine cannot answer is refused.
What Argent does with OPA
Open Policy Agent is the engine behind Argent's policy decisions, unless a tenant connects its own engine. Nothing to install and nothing to connect.
OPA is a CNCF project that moved to the Graduated maturity level on January 29, 2021 (CNCF). Its rules are written in Rego, a language many security teams already read.
NIST's zero trust architecture, SP 800-207, puts it plainly: "Authentication and authorization (both subject and device) are discrete functions performed before a session to an enterprise resource is established." (NIST SP 800-207). Argent applies that to AI agents on every request: a policy engine decides, not the agent.
- One engine for every kind of traffic. The same evaluator decides model calls (
llm), tool calls (mcp), agent-to-agent calls (a2a), agent actions such as shell commands and file writes (action), and requests to use a stored credential (credbroker). - Default deny for models. Out of the box, a model call is allowed only when a model allowlist names that model. With no allowlist, the answer is
llm.deny.no_allowlist. - Default deny for agent actions. Blocked content, unmeasured content and unknown action kinds are refused first; anything no rule allows is refused as
action.deny.no_rule. - Fail closed. A missing rule, an evaluation error, a crash or a timeout is a deny, never an allow.
- No restart. Decisions are cached against the policy version, and every gateway reloads when a rule changes.
How it runs
OPA runs inside Argent's gateway, in the same process, so a decision never leaves your deployment.
| Where | In-process in the gateway and the control plane |
| Rules | The shipped Rego rulebook: base path deny list, models, MCP, MCP arguments, agent-to-agent, agent actions and the credential broker |
| Delivery | From disk, or as a digest-pinned bundle from an OCI registry, with a size cap that stops oversized zip-bomb bundles |
| Your edits | Stored in Argent's database and laid over the shipped bundle; if that read fails, the shipped rules are served, never no rules |
| Changes | Every gateway reloads on a policy change and reports the revision it runs, so a stale gateway is visible |
| Air-gapped | Bundles can come from a local registry, so nothing is fetched from the internet |
The engine switch accepts opa or the deprecated stub mode, and any other value refuses to start. A typo can never quietly weaken authorization.
What you configure
Everything is in the console under Policy. No YAML, no restart.
- Approve models. Set which models each agent, person, department or the whole tenant may use on the Model access page. See Model Control.
- Read and edit the rules. The Rego editor shows each shipped module with its history. No edit is saved unless it parses and the whole bundle still compiles with it.
- Test before you ship. The policy simulator runs a sample model, MCP or agent-to-agent request through the live engine and writes a record marked simulated. Impact replay runs a window of past decisions under the live and the proposed rules and reports what would change.
- Watch a module first. A SuperAdmin can set a shipped module to watch, so its refusals are recorded as would-have-refused instead of enforced. The base module can never be watched. See Watch Before You Enforce.
- Keep a change window. Changes outside the tenant's weekly window are flagged, and a change can carry a note from its author, kept in that policy's history.
Changes need proof and sometimes two people
A policy engine is only as safe as the way its rules change.
- Two keys on the base rules. A save, rollback or restore of
argent/base.regowaits until a second SuperAdmin confirms it within 24 hours. Confirming your own change is refused. - History and undo. Every policy has its own history, undo and revert. A revert is itself recorded and can be undone.
- Archive, not delete. Archiving a policy keeps a snapshot, stops enforcing it and signs the event in one step.
- Which rule said no. Every refusal carries the rule id that made it, such as
llm.deny.model, on Argent's signed record.
Argent as a decision point for your own tools
The same engine can answer other software, over an open standard.
Argent's gateway serves the OpenID AuthZEN 1.0 evaluation and batch evaluations calls at /_argent/access/v1/. The OpenID Foundation announced Authorization API 1.0 as a Final Specification on January 12, 2026 (OpenID Foundation; specification). Hooks, SDKs and third-party agents can ask Argent whether an action is allowed.
- Off until you turn it on. The decision API is off for a tenant until an administrator switches it on, and it has its own per-minute rate limit.
Where this fits
Pages that go deeper.
- Your own engine instead? A tenant can have the gateway ask an engine it already runs about its agents' traffic: OPA server, Cerbos, AuthZed SpiceDB or Topaz.
- Documents too: retrieval-time authorization for RAG runs on OpenFGA.
- Capabilities: Model Control, Agent Permissions, Watch Before You Enforce and all capabilities.
- Goals: model control and routing, agent identity and ownership, audit and accountability and all goals.
- Use cases: controlling agent tool access, securing AI coding agents and AI in regulated and air-gapped environments.
- Where it runs: air-gapped on-premises, private cloud or your public cloud VPC. Compare them on deployment options.
- Industries: Financial Services, Insurance, Healthcare, Public Sector and all industries.
- More integrations: every system Argent connects to, grouped by type, on the integrations page.
Want this running on your own use cases?
All downloads