What’s next for Microsoft’s Copilot? AI agents that do tasks for you

You’ve probably barely become used to interacting with ChatGPT, Copilot, and other AIs. Whether you like it or not, Microsoft will soon start guiding you towards the next generation of AI: ordering agents to autonomously pursue tasks for you, particularly in the business space.

Today, at Microsoft Build, the company begins rolling out tools to consumers and business users that show off its vision. Microsoft 365 Copilot Wave 2, for example, is Microsoft’s name for an improved Microsoft 365 application, a hub where your chats, notebooks, and agents all intersect. It will include Copilot Search, too. Microsoft wants those business users to also begin training their agents and models on company data so that a legal agent, for example, can quickly come up with a merger proposal or disclaimer form.

Other work that Microsoft is doing is laying the groundwork for a future that isn’t here yet, but you’ll probably want to know about: what Microsoft calls a Model Context Protocol will be a framework for AI agents to work with and control native Windows apps—part of what Microsoft originally promised with Copilot but never really delivered. And while Microsoft is pushing what it calls Windows AI Foundry and Windows AI Foundry Local toward developers, you might benefit, too—they’re designed to allow LLMs to run locally, with Microsoft handling all the deep thinking about bits of code you’ll need and how they’ll be optimized for the hardware that’s already on your PC.

Microsoft seems to believe that any ground it lost to OpenAI, Meta, or Anthropic can be quickly made up by pushing AI across its workplace and yours. If it’s scary, just think: Microsoft just went through a round of layoffs that apparently were used to replace some of programmers with AI.

What’s next for you in Microsoft’s AI

So what’s an agent? If you’ve interacted with Copilot, ChatGPT, Claude, or others, you might have delved into what some call “Deep Research”: you ask a question, the AI formulates a plan, and then it goes off and figures out what you’ll need. Agents are designed even more autonomous so that you’ll be able to assign them a task, and then they’ll keep at it until it’s done—potentially even repeating the task. That’s the hypothetical approach, anyway.

Microsoft 365 Copilot Wave 2 will probably have the most immediate impact on your workday, provided that you work for a company that subscribes to Microsoft 365. The app itself is designed for human-agent interaction, and Microsoft showed off a menu with an “agent” category. Two of those, Researcher and Analyst, will be included. They’re rolling out to customers by what Microsoft calls the “Frontier Program,” which is basically a Windows Insider program for AI.

Microsoft has two demonstration videos that explain Researcher and Analyst. Researcher is like a supercharged version of Deep Research, but instead of making a plan and checking it with you, it might bounce questions or ideas off you to get things going.

Analyst feels like something that normal people could use—or at least those who haven’t learned good data practices.

In the Analyst video below, you’re essentially able to ask Copilot to format your data in a way that makes sense, like to whip it up for a meeting. Another, separate feature that Microsoft is working on is in PowerBI, where you’ll have an opportunity to “query” or ask questions of your data. At this point, Microsoft is excellent at taking related functions and sprinkling them across their various apps where they make sense.

Microsoft already offers what it calls Copilot Notebooks, where you can query Copilot and then turn the output into an interactive document. Microsoft says you can do this on your phone, and when you’re done, you can save the notebook as a “legacy” file format like Word for sharing.

Microsoft also said that Microsoft plans to roll out Copilot Search and Copilot Memory in June. We’ve already seen Copilot Search, which is live now; the AI-powered search slurps up the recommendations from your favorite content creators and journalists and then presents them in Microsoft’s voice, with small links.

Copilot Memory is more than just a history of your search. It’s a reminder of how you solved your problem. Currently, engaging with AI is a one-time, transactional process. Microsoft wants more of a history where it remembers what Copilot did and how it got there.

“So you interact with an agent, maybe it recalls past interactions,” Kevin Scott, Microsoft’s chief technical officer, said Sunday night. “It almost certainly doesn’t remember its scratch work over time, like the way that we would, like we solve a problem once, and then we sort of write it down somewhere on a piece of paper, store it on disk, or remember it. So memory is one of these problems that is really going to be important for us to solve, and it needs to be a form of agentic memory that probably more mirrors what happens with biological memory.”

Microsoft is also going to beef up Copilot’s Create function with OpenAI GPT-4o image generation, which OpenAI believes excels at generating text. So if you want a raccoon writing lines on a blackboard, the words will (hopefully) make sense.

Another feature, Microsoft 365 Copilot Tuning, might not be totally aligned with Microsoft’s Wave 2 efforts. But Microsoft is trying to allow Copilot to train itself on as much corporate data as you (or your company) will allow it to take over menial tasks.

“Users will be able to automate repetitive tasks using Computer Using Agent (CUA) technology for tasks such as data transfer, document processing, market research, and compliance monitoring,” Microsoft says.

Finally, Microsoft is taking its business browser, Edge, and improving it at understanding files shared over the Web, like Word, Excel, and PowerPoint.

Laying out an AI agent future

Microsoft is also setting the stage for future improvements, which aren’t here yet. Remember how Microsoft promised that Copilot would allow you to control your PC? That really never happened, although Microsoft’s promised controls of the Settings menu are an intermediary step.

According to Microsoft, the company is “evolving Windows for the agentic future” with a technology called Model Context Protocol, or MCP. (It’s possible Microsoft will rename this when they realize the similarities with Tron‘s Master Control Program.) Microsoft is also working to implement “app actions” within Windows, too.

“MCP integration with Windows will offer a standardized framework for AI agents to connect with native Windows apps, enabling apps to participate seamlessly in agentic interactions,” Microsoft said in a blog post authored by Pavan Davuluri, the corporate vice president in charge of Windows and devices at Microsoft. “Windows apps can expose specific functionality to augment the skills and capabilities of agents installed locally on a Windows PC.”

For developers, Microsoft is building Windows AI Foundry. It’s not obvious what AI Foundry does, but it appears to take AI models and model catalogs like Ollama and bring them within Windows so that developers can quickly try out new models. A complementary AI Foundry Local service will allow those models to run locally on a PC rather than in the cloud.

“In preview, Foundry Local will make it easy to run AI models, tools, and agents directly on devices, whether Windows 11 or MacOS,” Microsoft said. “Foundry Local will be included in Windows AI Foundry and will deliver best-in-class AI capabilities on Windows with excellent cross-silicon performance and availability on millions of Windows devices.”

“During preview, developers can access Foundry Local by installing from WinGet (winget install Microsoft.FoundryLocal) and the Foundry Local CLI to browse, download, and test models,” Microsoft added. “Foundry Local will automatically detect device hardware (CPU, GPU, and NPU) and list compatible models for developers to try.”

If you’ve ever tried playing around with AI yourself, you probably know that trying to figure out what model your PC can run, then downloading it and whatever dependencies it needs, and then trying to update it—it’s all a real pain. Foundry AI and Foundry Local may be aimed at developers, but this might be a tool for enthusiasts to keep an eye on as Microsoft moves ahead into its agentic future.