Governing Agent Authority

AI agent governance for identity, permission and proof in enterprise deployments

Picture of Darshan Patil

Darshan Patil

Chief Technology Officer, Futran Solutions

Futran_GuardRail_Governing_Agent_Authority

Get the white paper

The full 13 page PDF, with every source linked. A copy is also sent to your inbox.

In brief

AI agent governance is usually framed as a model safety problem. This paper argues the exposure sits elsewhere: an enterprise agent acts through credentials, and those credentials were granted to a system, not to a decision.

It makes three arguments. The control that matters for agents is delegated authority: which agent acted, on whose behalf, under which policy. That control cannot live inside the model, because published research shows adaptive attacks defeat model level defences at high rates. And evidence that authority was governed has to be produced at the moment of action, because it cannot be reconstructed afterwards.

Contents

Executive summary

The four tests, in brief

The paper sets out tests rather than a checklist. Each one has a pass condition a skeptical auditor would accept.

Test 1

Attribution

Can the organization name the agent, the person or process that delegated the task, and the policy version behind a randomly chosen action?

Test 2

Refusal

Is there a record of an action the control denied before it executed? A control that has never refused anything has not shown that it operates.

Test 3

Completeness

Can the decision record be shown complete and unaltered, with no route from an agent to a tool that bypasses it?

Test 4

Interruption

Can one agent be stopped without stopping the others, with a record of when and by whom?

Frequently asked

What is AI agent governance?

AI agent governance is the set of controls that decide what an AI agent may do, under whose identity, and with what evidence. This paper treats it as a question of delegated authority: which agent acted, on whose behalf, under which policy, with a record that shows it.

Increasingly, within their own platforms. Policy in Amazon Bedrock AgentCore evaluates agent tool calls against Cedar policies, and Microsoft Foundry adds tool call guardrails in preview. The paper explains the gaps that remain for most enterprises: mixed estates, fragmented evidence, and policy that no platform decides.

The paper sets out four tests rather than a checklist: attribution, refusal, completeness and interruption. Each has a pass condition an auditor would accept, and each can be run against agents already in production.

Read next

PRISM Agent Readiness Assessment

A fixed fee engagement that inventories the agents in scope and the credentials they use, runs the four tests against a sample of real agent actions, and reports which actions are governed, which are merely permitted, and what evidence exists for each.

This white paper is for general information only and is not legal advice. Futran Solutions is not a law firm, does not issue certifications or audit opinions, and does not replace the judgment of an organization’s counsel, auditors or regulators.