Microsoft Tests A Persistent Copilot Agent, Aiming To Blend Local Control With Enterprise Safeguards

Microsoft is quietly testing a version of its 365 Copilot that would act more like the local, autonomous agents built by open-source projects such as OpenClaw, according to people familiar with the effort. The planned capability would let Copilot carry out multi-step work on behalf of users over extended periods, a shift that could give office workers and IT teams a tool that does more than answer questions — while raising fresh security and governance questions.

OpenClaw, an open-source toolkit that spins up agents on a user’s machine to perform tasks automatically, prompted a debate after several high-profile misuse concerns. Microsoft’s prototype would aim to offer similar continuous assistance but wrapped in enterprise controls — identity boundaries, auditing and other limits — that large organizations demand. A company representative declined to outline technical details but said Microsoft is testing features that prioritize compliance and reduce friction for everyday tasks.

The company already sells several assistant-style features inside Microsoft 365, though most operate in the cloud. In recent months Microsoft introduced Copilot Cowork, which can act inside Office apps instead of only replying in a chat pane, using a personalization engine the firm calls Work IQ; and Copilot Tasks, a preview feature positioned more toward prosumers for automating things from inbox triage to scheduling. Two of the three Copilot-like offerings available today run on Microsoft’s servers rather than on local machines.

Microsoft has also broadened the set of models available to its customers, offering Anthropic’s Claude as an option for some Copilot experiences after a partnership announced late last year. OpenClaw projects likewise often use Claude, a reason some users of the open-source tools prefer that model. But unlike OpenClaw, which many hobbyists run on compact Mac hardware, Microsoft’s commercial customers often require tight controls around data residency and remote management.

Security experts say a locally running Copilot-style agent would force a rethinking of enterprise policies. “Businesses will want role-based limits, detailed logs and the ability to revoke actions if an autonomous assistant goes off course,” said an enterprise security consultant who asked not to be named. They added that firms will also pressure vendors to certify how agents behave on different operating systems and hardware.

There are operational incentives for Microsoft to move in this direction beyond matching OpenClaw’s functionality. Some IT buyers prefer tools that minimize cloud round trips for latency-sensitive work or that give admins greater oversight of automation. At the same time, a local-agent option could complicate Microsoft’s product messaging, creating an additional layer for deployment and support teams to handle.

Microsoft is expected to discuss updates to Copilot-related products at its Build developer conference in June, though the company hasn’t confirmed a formal launch. A spokesperson told this publication that the company routinely pilots new orchestration and autonomy features across consumer and enterprise products, and that any wider release would come with governance and security protections tailored to customers’ needs.