GenAI & Agents

AI Agent Controls

A written scope for each AI agent, checked on every action it takes.

GIST stepTrust
Download the PDF

Get the AI Agent Controls overview

A short PDF to share with your team. It covers what is in the package, how it works, and a worked example.

The problem

Agents take real actions. Security needs to see the limits.

An agent that sends email, updates records, or queues payments needs the controls you would give a new hire with system access. Most agent projects add them late, if at all. We run these same controls on the agents on our own team.

What's included

A written scope, enforced on every action.

01

Scope files

One file per agent listing its allowed tools, data, limits, approvers, and a review date. Scope files are reviewed like code and stored where the agent cannot edit them.

02

The guard

Wraps the tool functions an agent calls, so it works with any framework. An operator CLI lists pending approvals, approves or denies them, stops and resumes agents, and verifies the log.

03

Human approval

High-impact actions wait for a named approver who is not the agent’s owner. Requests go to a file queue by default, or to Teams, ServiceNow, Jira, or email.

04

Audit log and switch-off

A hash-chained JSON log of every action, block, and approval shows any edit, removal, or reordering, and is copied to Azure Monitor or your SIEM. A stop takes effect on the agent’s next tool call.

How it works

What every tool call is checked for.

  1. Switched off?

    A stop file for this agent or for all agents, an environment switch, or a paused scope.

  2. Prohibited?

    A baseline list no scope can remove, such as changing its own permissions or turning off logging.

  3. In scope?

    The tool must be listed and every data resource allowed. A denial always wins.

  4. Within limits?

    Actions and cost, per run and per day.

  5. Approved?

    High-risk tools wait for a listed approver who is not the owner. No answer counts as a no.

  6. Logged either way

    Success, failure, cost, and duration are recorded. A blocked call goes back to the model with the reason.

Worked example

An accounts-payable agent that triages invoices.

The agent checks invoices against purchase orders. The demo runs through every control.

What happens
  • Mismatched invoices are flagged, and matched ones wait for payment approval
  • Reading supplier bank details is blocked
  • Releasing a payment is prohibited
  • Changing its own permissions is refused
  • Approval from someone off the list does not count
  • After switch-off, every call is blocked
Result

Every event is in the audit log, and the hash chain proves nothing was changed.

Where it fits

What it runs on, and where it stops.

Works with

  • LangChain, Semantic Kernel, AutoGen, the OpenAI and Anthropic SDKs, or your own code
  • Azure managed identities and Azure Monitor
  • Teams, ServiceNow, Jira, or email for approvals

Assumptions and limits

  • It controls the tools you wrap, so we pair it with platform permissions
  • Cost limits depend on the costs you feed in from your model provider
  • Agents running on several machines need shared state storage
Ask Dex

Ask about the AI Agent Controls

Dex answers from our own pages and links to where it found the answer.

Is security holding up an agent launch?

Bring the list of tools your agent calls. We will draft its scope file with you.