Microsoft Copilot Cowork Launches in Early Access With Critique and Model Council

Microsoft has opened early access to Copilot Cowork through its Frontier program, bringing the company’s most ambitious agentic AI feature yet to Microsoft 365 users. The rollout is accompanied by two significant upgrades to Researcher — the deep research tool inside Copilot — that introduce multi-model verification and side-by-side model comparison directly into enterprise workflows.

What Copilot Cowork Actually Does

Built on top of Anthropic’s Claude Cowork, Copilot Cowork is an agentic AI tool designed for long, multi-step tasks inside the Microsoft 365 ecosystem. Unlike conventional Copilot interactions — which are largely single-turn — Cowork takes an outcome-oriented approach. You describe the result you want, and the system builds a plan, executes it step by step, and surfaces its progress as it works.

Users can intervene and redirect at any point, making it a supervised autonomous workflow rather than a hands-off automation. The tool handles both one-off requests and repeating workflow templates, such as monthly budget reviews or recurring report generation — tasks that currently consume significant knowledge worker time with low creative return.

According to Microsoft’s announcement, Cowork is positioned as the connective tissue between individual Copilot interactions and full enterprise workflow automation — a colleague-like layer that can operate across documents, emails, and workflows without requiring the user to manage each step manually.

Two New Researcher Features: Critique and Model Council

The Cowork launch is paired with a meaningful upgrade to Researcher, Microsoft’s deep research feature inside Copilot.

Critique: A Dual-Model Review Pipeline

The first addition is Critique — a setup where two AI models work sequentially on the same research task. OpenAI’s GPT generates the initial response; Anthropic’s Claude then reviews it for accuracy, completeness, and quality before the result surfaces to the user.

Reuters reports that Microsoft plans to make this relationship bidirectional in future releases — meaning Claude drafts could eventually be reviewed by GPT in return, creating a genuine mutual verification loop rather than a one-directional quality filter. Microsoft says Critique improved Researcher’s score by 13.8% on the DRACO benchmark, the industry standard measure for deep research accuracy and quality.

Model Council: Side-by-Side Model Comparison

The second addition is Model Council — a feature that lets users pull responses from multiple AI models and compare them simultaneously in a single interface. Users can see immediately where models agree, where they diverge, and what each brings distinctively to a given question.

For enterprise users dealing with high-stakes research, strategic decisions, or policy analysis, Model Council introduces a layer of epistemic triangulation that single-model outputs cannot provide. It also positions Microsoft as a model-agnostic orchestration layer rather than a single-vendor AI dependency — a framing that matters considerably for enterprise procurement conversations.

Wave 3 and the Shift From Assistant to Agent

Microsoft frames both launches as part of Wave 3 of Microsoft 365 Copilot — a deliberate push to move AI in the workplace from an experimentation tool to one that actively executes work. Wave 3’s defining characteristic is agency: the ability to plan, act, verify, and complete tasks across an extended workflow rather than responding to isolated prompts.

The dual-model Critique architecture and the Cowork agentic framework both reflect the same underlying conviction: that the next phase of enterprise AI adoption will be won by systems that reduce cognitive load at the workflow level, not just the sentence level.

Copilot Cowork is available now in early access through Microsoft’s Frontier program for eligible Microsoft 365 subscribers.