AI agents on Windows: opportunities and controls for companies

Microsoft no longer wants to be just the AI “integrator”

For a long time, when we talked about Microsoft and artificial intelligence, the conversation revolved around OpenAI, Copilot and the integration of AI into products such as Office, GitHub or Windows.

But at Build 2026 the message changed quite a bit: Microsoft wants more control over the entire chain.

It doesn't just want to put AI inside its products. It wants its own models, its own infrastructure, its own system to train them and its own environment to run agents safely. Microsoft announced a family of seven in-house models called MAI, focused on reasoning, code, image, voice and transcription.

Put simply: Microsoft is no longer playing only as a “model customer” and is moving as a full-stack maker.

What the MAI models are

The MAI models are Microsoft AI's new family of models.

Among the main announcements is MAI-Thinking-1, a reasoning model designed for complex problems, math and software engineering tasks. Microsoft also introduced MAI-Code-1-Flash, a lighter, more efficient model for programming, integrated with GitHub Copilot, VS Code and the Microsoft ecosystem.

There are also models for image, transcription and voice:

The important part isn't just that they exist. The important part is that Microsoft says it trained them from scratch, with clean data and without relying on distillation from third-party models.

For companies, that matters a lot.

Because when a large company adopts AI, it doesn't only ask “how smart is it?” It also asks: “where did it come from?”, “what data was it trained on?”, “can I audit it?”, “can I control it?” and “how much will it cost me to run?”

The big one: OpenClaw on Windows

The other interesting part of the event was OpenClaw for Windows.

Microsoft announced that OpenClaw can now run natively on Windows using Microsoft Execution Containers, or MXC. MXC is a layer that lets you define what an agent can touch: files, network, permissions and system resources.

The simple analogy would be this:

An AI agent without limits is like handing the keys to your house to someone you barely know. MXC tries to give that agent an access card: it can enter certain doors, but not the whole house.

And that is key, because agents aren't simple chatbots.

A chatbot answers you.
An agent can act.

It can open files, read information, run tasks, connect to services, automate workflows and make decisions within an environment. That is powerful, but also delicate.

Windows wants to be the safe place to run agents

Microsoft also talked about identity, containment and enterprise management for agents on Windows. The idea is that an agent isn't “something invisible” running on your machine, but an entity with rules, permissions and oversight.

This makes sense for companies.

You can't bring autonomous agents into an organization without control. It would be like hiring a new person and giving them access to every folder, system, email and database… but with no badge, no log and no boss.

That is why Windows is taking an interesting position: it no longer wants to run only apps, it now wants to run agents.

The essentials

Microsoft Build 2026 left three clear signals:

1. Microsoft wants its own models, not to depend entirely on third parties.
2. Windows is getting ready for local agents, not just traditional applications.
3. Security will be a central part of the future of AI, because an agent that runs tasks needs real limits.

For devs, this means the near future won't be only about learning to use prompts.

It will also be about understanding how models, permissions, files, tools, identity, observability and security connect.

What the reader takes away

AI is moving from “answering nicely” to “doing things.”

And when an AI starts doing things, the problem changes.

It's no longer enough to ask whether the model is good. We also need to know where it runs, what permissions it has, how it is audited, what data it touches and how it is stopped if something goes wrong.

Microsoft is betting that Windows will be that controlled place to run agents. And if it succeeds, it can turn Windows into more than an operating system: a work platform for AI agents.

The change is not minor.

Because the next leap in AI won't be just talking to it.
It will be letting it work… but with clear rules.