Microsoft tests an OpenClaw-style AI agent for Copilot with a focus on security

Microsoft is testing a new agent for Microsoft 365 Copilot that mirrors OpenClaw's logic but is designed for enterprise use with stricter security controls. The company has not yet disclosed its architecture, but has already confirmed experiments with features that will allow Copilot to not only respond to requests but independently execute long chains of actions.

What Microsoft is preparing

The discussion centers on expanding Microsoft 365 Copilot with agent capabilities that can take on real user tasks. According to reports, the new tool should be able to work continuously, rather than launching only on command within a specific chat window. The idea is for Copilot to accompany processes in the background, make decisions within defined rules, and drive complex multi-step tasks to completion.

It remains unclear whether this agent will run locally on a device, like OpenClaw, or whether Microsoft will limit itself to bringing some of its approaches into Copilot's cloud infrastructure. For enterprise customers, this is an important distinction: local mode provides greater flexibility and can accelerate certain scenarios, but significantly increases requirements for isolation, access management, and auditing of agent actions.

How it differs

The main difference from OpenClaw lies not in the concept of an autonomous assistant itself, but in how Microsoft is trying to package it for business. OpenClaw gained popularity precisely because it can act on behalf of a user, but along with that quickly acquired a reputation as a risky tool due to broad permissions and weak protective boundaries. If Microsoft brings the project to release, the bet will be on enterprise control mechanisms:

• stricter security policies and governance
• built-in rules defining what actions the agent is permitted to take
• better integration with enterprise identity and access management
• ability to track what the agent did and why
• emphasis on trust and reducing daily friction for employees

This emphasis is understandable: companies are willing to pay for autonomy only when they can constrain it within compliance and internal policy frameworks. For Microsoft, this is a particularly sensitive topic because an agent acting on behalf of an employee gains access not to abstract data, but to email, documents, calendars, and business processes. Therefore, security here appears not as an optional add-on but as a core product requirement.

"We are constantly experimenting," said a

Microsoft representative.

Why the company needs a new agent

Microsoft already has several agent products, so the new project looks not like starting from scratch but like another layer on top of an already established lineup. In March, the company introduced Copilot Cowork — a tool capable of acting within Microsoft 365 applications, not just responding in a side panel. It runs on cloud infrastructure and uses a personalization layer that Microsoft calls Work IQ.

Earlier, in February, Copilot Tasks appeared. This agent is designed to perform tasks like email triage, travel planning, and meeting organization. But it too operates from the cloud.

Against this backdrop, the new project can address a different scenario: long multi-step processes that run almost continuously and require greater autonomy than a typical chatbot or one-time workflow.

There is another motive: positioning. For Copilot, it is important to remain the central entry point for work tasks while the market increasingly discusses local agents capable of using interfaces almost like humans. The article particularly notes that OpenClaw can run on Windows, but among active users, Mac Mini is especially popular. For Microsoft, this is a clear signal not to cede this new category of AI tools to another platform. The product reveal is expected at the Build conference in June 2026.

What it means

The race for AI agents is rapidly shifting from laboratory demos to enterprise products with clear constraints, action logs, and access controls. If Microsoft truly builds such a mode into Copilot, companies will get not just another assistant, but a more autonomous automation layer for long office processes — but only if security proves to be not in the presentation, but in the actual product architecture.