The next frontier in AI and productivity -- reasoning agents at work

As Satya Nadella announced in the Microsoft AI tour earlier this week, we have been working on a new category of agents for work as part of Microsoft 365 Copilot – Analyst and Researcher.

I want to share more context about how we are building these products at Microsoft .

Evolution of AI from Tools to Agents

As model capabilities advance, AI products are evolving from specific tools to sophisticated agents. We are building these agents around key product dimensions – the ability to run independently while supporting human in the loop, the capability to verifiably solve complex tasks, and the power to learn to use the right tools (or even other agents) to make progress on assigned tasks.

As we started building, we identified three challenges to build products that can be most useful for our customers.

1.    How do we build agents that reason over a large amount of work data in a secure and compliant way?

Information in the world at large is useful, no doubt. But as an information worker, you need AI to be able to reason over your work world, structured data as well as the files, people, meetings, emails that are most relevant to your role and daily tasks.

2.     How do we build agents that work where and how you work?

Our approach is to build and support an agentic system – where apps, assistants and agents all work together - as part of M365 Copilot. We are building these agents to be easily integrated into your current workflows and the Microsoft 365 apps you already use.

3.     How do we build agents that let you work with others?

Modern work is deeply collaborative and any products we build need to let you share the insights, analysis and work in progress with others through documents, sheets and decks that you then work together with others.

Early Results

We are rolling out Analyst and Researcher as part of "Frontier" program -- an early-access program for cutting-edge capabilities and features. We are super encouraged by the early evals and by the model-forward product making approach we are taking here -- developing unique capabilities (e.g. reinforcement learning and dynamic code execution in the case of Analyst agent or enabling safe and compliant tool calling from within the models in the case of Researcher agent), and building for depth of work use-cases.

What’s Next?

This is just the first step in our long-term vision of building an agentic system of apps, assistants, agents that works for you at work. We will continue to build new model capabilities, combine them with the depth of workloads that we can uniquely serve and enable these agents to work in tandem and as part of M365 Copilot and the apps. For example, the Researcher agent will soon be able to tap into the sales agent to give you the best briefing document ahead of a prospective customer meeting.

For a peek under the hood on how we built and evaluated these agents, check out the Researcher and Analyst Microsoft technical community blogs. To learn more about the announcement, check out the Microsoft 365 Community blog.

Excited to shape the next frontier of AI and productivity with agents for work and looking forward to learning and building together with our customers!