Briefing
Reference 6 · Primer · About 5 minutes
Download the MP3 · Linda, an AI voice, reads a script written from the sources below.
Written from public sources and checked on 10 October 2026. The toolkit is public preview software and its details change often. Check the current documentation before relying on it.
Agents plan, use tools and act over many steps. The governance question is who owns them, what they may reach, and how they are stopped.
These are Agile Insights AI Strategy & Governance offerings.
Sign up for our podcast at agile-insights.com.au.
Linda The Microsoft Agent Governance Toolkit is a free tool that stops an AI agent doing something it should not do, at the moment it tries. By the end of this episode, you will know what it is, what it covers, and what it leaves squarely with you. This is as at October 2026.
Linda Start with what it is. Microsoft published it in April 2026, announced on the Microsoft Open Source Blog. It is open source software, released under the MIT licence, which means anyone can use and change it at no cost. It is not a standard, and it is not a certification. It is code that developers place between an AI agent and the tools the agent uses. It is a toolkit, not a service Microsoft runs for you, and it works outside the Microsoft world too.
Linda Here is the problem it solves. Many organisations are now giving agents real access: email, databases, payment systems. Access permissions control which systems an agent can reach. They do not control what the agent does once it is inside. Imagine an agent allowed to read customer records that decides, on its own, to send them to an outside address. The permissions allowed the read. Nothing checked the intent. This toolkit is built to check that, before the action happens.
Linda Who uses it? Mostly engineering and security teams, who install it. But executives meet it in other ways. A risk lead asks, how do we know this agent stayed inside its limits? An auditor asks for evidence. A board asks who is accountable when an agent acts. The toolkit produces some of the evidence for those conversations. It does not produce the answers.
Linda Now, what is in it. Four ideas carry it. The first is policy enforcement. Every action an agent attempts is checked against written rules, and the answer is allow, warn, block, send to a human, or modify. If the check itself fails, the action is blocked. The second is identity. Each agent gets its own verifiable identity, so you can tell which agent did what, instead of five agents sharing one login. The third is isolation, which means limiting what an agent can touch, with a kill switch to stop it. The fourth is evidence. It keeps a tamper evident record of each decision, so you can show later what was asked and why it was allowed.
Linda How does it connect to the rest of the Microsoft stack? Think of two layers. Microsoft Agent 365, which became generally available on the first of May 2026, is the tenant wide control plane. It keeps a registry of agents, and works with Microsoft Entra Agent ID, which gives each agent a directory identity and a named human sponsor. Microsoft Purview adds data protection and audit. Those tools look at agents across the whole organisation. The toolkit sits closer to the action, inside the agent's own runtime, and decides before each step. It plugs into Microsoft Agent Framework, and the project publishes a guide for deploying it with Foundry Agent Service. Its documentation also describes a bridge to Entra Agent ID, though that is a project tutorial, not a Microsoft product feature.
Linda It also maps to the OWASP list of the ten biggest risks for agentic applications, published in December 2025. The project's own page says it covers all ten. Its repository badge says seven fully and three partly. Either way, that is the project grading itself. It states plainly that this is not a third party audit.
Linda Now the limits, because they matter most. The toolkit governs actions. It does not judge whether the agent reasoned well. It records attempts, not outcomes. It does not govern the documents an agent reads. And if no policies are loaded, it allows everything by default. A tool that looks switched on, but is not configured, is worse than none. None of this replaces an owner for each agent, decisions about scope, a risk register, or independent assurance.
Linda Three first steps. One, make sure you have a register of agents, each with a named accountable owner. Two, decide which agent actions need a human to approve them. Three, ask your technology team to pilot the toolkit on one low risk agent, and test that it blocks what it should.
Linda For primary sources, start with the project's GitHub repository and documentation site, the Microsoft Open Source Blog announcement, Microsoft Learn on Agent 365 and Entra Agent ID, and the OWASP agentic top ten. Links are in the show notes.
Linda On currency. The toolkit is in public preview. The project says it is production quality, but it warns of breaking changes before general availability, and it has given no date. The latest tagged release is version 4.1.0, from the ninth of June 2026, though the code was still changing this week. The project says it is moving towards shared governance across several organisations. Package names have changed more than once, so check the current documentation before relying on any detail I have given.
Linda To recap, this toolkit checks and records what an agent does, but people still have to decide what agents should do. If you'd like to know more, sign up for our podcast at agile-insights.com.au.
Every episode tagged #microsoft
All of the reference series (16) · Follow AI Governance Reference in Apple Podcasts or any podcast app
Everything from Agile Insights: briefings, Making You Think and the reference series