On Thursday, Microsoft introduced Copilot Tasks — a new feature that uses cloud computing and a browser to perform work on behalf of the user.

This is the most concrete step the company has taken towards turning its AI assistant into an autonomous agent capable of executing multi-step workflows without constant human oversight.

A feature that is rolling out in a limited research preview with a public waiting list
allows users to describe their goal in natural language. Copilot then creates a step-by-step plan, executes actions in applications and web services in the background, and reports when the work is completed or when approval is needed.

From «Help me» to «Do this for me»

Since its launch, Copilot has primarily operated as a communication tool — summarizing documents, composing emails, and answering questions. Copilot Tasks represents a conscious shift from this chat-oriented model to what Microsoft describes as a task list that «executes itself».

«Many AIs still require too much configuration and too much blind trust,» wrote Yusuf Mehdi, head of Microsoft, in a LinkedIn post announcing the launch. «Copilot Tasks is designed so that you can delegate work in plain language, see the plan, maintain control, and get a finished result.» Mehdi added that the company is «opening Copilot Tasks as a research preview for a small group of interested parties to try, test, and help make it better», with
expanding access through a waiting list.

Tasks can be executed once, on a schedule, or until a specific condition is met. According to Microsoft, early use cases include preparing daily email summaries with draft responses, monitoring rental listings, and booking viewings, compiling meeting briefings from
calendar data, transforming a syllabus into a lesson plan, comparing contractor bids, and tracking hotel prices for automatic rebooking when prices drop.

Limitations and consent

The system operates in a controlled cloud environment with its own browser instance, meaning it does not rely on the user's local computer. Microsoft emphasizes that Copilot Tasks is not fully autonomous — it requests user consent before performing actions related to spending money, sending messages, or confirming bookings. Users
can pause or cancel workflows at any time.

This consent model will be closely examined as the feature scales.
Security researchers have previously noted that autonomous browsing combined with execution — clicking buttons, filling out forms, submitting data — creates a vulnerability for prompt injection and malicious web content. Microsoft’s own documentation for browser automation tools acknowledges these risks.

Competitive landscape

Copilot Tasks appears against the backdrop of an industry-wide movement towards agent-based AI. Operator from OpenAI, Nova Act from Amazon, Project Mariner from Google, and the computer usage feature for Claude from Anthropic — all aim to allow AI systems to navigate websites and perform tasks on behalf of users.

Microsoft's advantage lies in its extensive ecosystem —
Windows, Edge, Microsoft 365, and Azure, — which provides deep integration points that competitors lack. The feature builds on earlier capabilities of Copilot, including Copilot Actions, which could perform specific automated steps, and the Researcher and Analyst agents introduced in mid-2025.

Whether Copilot Tasks will fulfill its promise will depend on reliability. As stated in Microsoft's promotional video: «You describe the task, and Copilot takes care of the rest.» How well it manages the rest — and how comfortable users are allowing it to try — remains an open question.