Microsoft has introduced a new line of AI agent products for enterprise customers, extending Copilot from an AI assistant into a broader set of agentic capabilities. Announced as “Wave 3” of the Microsoft 365 Copilot rollout, the launch centers on a new agent called Copilot Cowork, built in collaboration with Anthropic and powered by its Claude models.
Copilot Cowork and the move toward agents
The reliance on Anthropic marks a notable shift for Microsoft, which has been closely associated with Anthropic’s rival OpenAI. Copilot Cowork is a cloud-based agent designed to plan, execute, and deliver complex, multi-step work from a single request, acting across applications such as Outlook, Teams, Excel, PowerPoint, and SharePoint. By drawing on a user’s work apps and context, Microsoft says it can produce more targeted, tailored results. The agent entered a research preview on March 9, 2026 with a limited set of customers, with broader availability through Microsoft’s Frontier program expected later in the month.
As part of the same release, Claude has also been made available in mainline Copilot chat through the Frontier program, alongside the latest OpenAI models, with Copilot selecting the appropriate model for a given task. Microsoft additionally outlined a new generation of agentic features for Word, Excel, PowerPoint, and Outlook, including an improved chat experience, the ability to create and refine artifacts, and tools that let employees build their own agents.
Enterprise packaging and governance
Microsoft is bundling these capabilities into a new enterprise suite, Microsoft 365 E7, which combines Microsoft 365 E5, Microsoft 365 Copilot, and Agent 365 into a single offering integrated with the company’s security stack. Microsoft also said its Agent 365 system, aimed at managing and governing agents, would become generally available on May 1. Governance featured prominently in the announcement: a senior Microsoft executive noted that rapid agent adoption signals clear customer value but warned that, without guardrails, that speed can create blind spots and security risks. A related look at the controls enterprises put around autonomous systems appears in this overview of enterprise AI governance.
What to watch
Several of these capabilities, including Copilot Cowork, are in research preview, so features, availability, and pricing may change before general release. Running an agent across email, chat, and documents raises real questions about data access, residency, and oversight, which is part of why governance tooling is being shipped alongside the agents. The practical value will depend on how reliably these agents complete multi-step tasks in real workflows and how well organizations can monitor and constrain what they do. As with any agentic system, output should be reviewed rather than trusted automatically, particularly where it touches sensitive business data.
