Microsoft introduced Researcher Agent with Computer Use in Microsoft 365 Copilot Chat. This extension enables secure interaction with public, gated, and interactive web content through a virtual computer. It supports deeper research, actionable workflows, and richer outputs grounded in organizational data and the web.
Timeline
Researcher Agent with Computer Use is available for licensed users with Copilot Frontier access.
How does this affect your Copilot users?
Researcher Agent with Computer Use builds on the existing Researcher capabilities. With Computer Use, the agent can securely interact with public, gated, and interactive web content through a virtual computer, enabling users to uncover deeper insights, take action, and generate richer reports grounded in both work data and web data.
Users can watch the agent search the web as it clicks, types, navigates, and completes tasks in a simulated environment.
Researcher Agent with Computer Use requires:
- A Microsoft 365 Copilot license.
- Access to the Researcher Agent.
- Access to the Copilot Frontier program. Note, the Frontier program is not for productive use cases.
- Access to the Computer Use feature. Computer Use is enabled by default.
- Computer Use is an optional extension. The original Researcher Agent remains available if the user is not enrolled in Frontier or if Computer Use is not enabled.
As Microsoft describes Computer Use:
When Computer Use is triggered, a sandbox environment dedicated to the conversation is created. This environment follows strict browser and network policies set by Microsoft. It’s isolated from both the intranet and the user’s device. Administrators can customize the sandbox by adding specific domains to allow or blocklists. Commands from the Researcher agent are sent securely. The sandbox environment doesn’t persist for long-term use and is ephemeral. It can only access the web with enforced policies, and no user credentials are stored or transferred in or out of the Sandbox. The virtual machine runs on Windows 365 for Agents and hosts containers running Linux.
Based on the scoping and configuration set by an administrator, Computer Use supports the generation of rich artifacts using advanced code generation, including presentations, spreadsheets, applications, and web content from public sources. These outputs can also incorporate business data from across the organization, such as meetings, files, and chats.
In my simulation, I asked the Researcher to search my Topedia Blog, the Icewolf Blog from Andres Bohren, and Microsoft Learn. The scope was limited to web content.
Note:
My simulation crashed several times. I was almost expecting to see a bluescreen from the sandbox. As usual, Copilot simply returned an error message. It’s a bit frustrating and unproductive when the Researcher starts working, and after 30 minutes, you end up with a message like the one below.
I then shortened and rewrote the prompt to avoid overloading the agent and its sandbox.
The agent searches my Topedia Blog…
…then the Icewolf Blog,…
…including reading posts from Icewolf.
At the end, the agent prepares a summary of the findings and offers to export the results as markdown.
I stored my simulated markdown export in my GitHub repository.
