DemoPricing

What is Pre-Action Authorization for AI Agents?

Pre-action authorization is checking each action an AI agent or automated system is about to take against the rules it must act within, before the action runs. The check answers for that one action, with its facts: allowed, denied, or held for a named person to approve. It authorizes the act, not the agent.

Why access control is not enough

Access control decides what an agent may reach: which tools, which APIs, which data. It does not decide whether this refund, this payment or this message is allowed under policy. An agent with access to a payments tool can still send a payment the policy forbids. Pre-action authorization sits at the moment of the call, and judges the action itself.

What a check needs

The rules, stated precisely enough to evaluate; the facts of the action (the amount, the customer, the country); and a verdict that says which rule decided. A check that cannot be answered because a fact is missing should hold the action and name the missing fact, never let it through by default.

How NOMOS does it

A policy or law is sealed as an artifact: each rule a condition and an outcome, citing the sentence it came from. Before acting, a system calls nomos.can() with the facts, or the nomos-guard MCP gateway checks every tool call, and gets AUTHORIZED, DENIED or ESCALATED with the rule. An escalated action waits until enough of the right people approve that exact call. Every answer is recorded, and the sealed artifact can be verified offline with its public key.

Before, not after

Monitoring and audit find a bad action after it happened. A check before the action prevents it, and the record of the check is the audit trail: what was asked, what was answered, under which version of the rules.

Example

Scenario

A customer-service agent can issue refunds through a tool. Policy says refunds over $500 need a manager.

Without governance

The agent has access to the refund tool, so a $2,000 refund goes through. The breach is found in the monthly review.

With NOMOS

The refund call is checked first: the rule for refunds over $500 escalates it to a manager. The agent waits; the manager approves that exact refund, once, and the decision is recorded with the rule.

Frequently asked questions

Is pre-action authorization the same as guardrails?+

Guardrails usually filter what a model says. Pre-action authorization judges what a system does: a specific action with specific facts, against stated rules, with a recorded verdict.

Does it work with MCP?+

Yes. nomos-guard sits between an agent and its MCP servers and checks every tool call against a sealed policy before forwarding it; denied calls are blocked with the rule that denied them.

What happens if the policy service is unreachable?+

The check fails closed: the action is held, not allowed. A sealed artifact can also be evaluated locally, so checks keep working without calling NOMOS.

Related concepts

Check an action before it runs

One call with the facts of the action: AUTHORIZED, DENIED or ESCALATED, with the rule that decided.

For developers →