Microsoft has introduced Copilot Cowork, an AI agent feature for its cloud productivity platform Microsoft 365, SiliconANGLE reported on March 30 local time.

Copilot Cowork helps AI agents handle long-running, multistep work that previously required constant human oversight. Users state the outcome they want, and the agent creates a plan and carries out the required tasks in sequence, focusing on “delegated” automation.

While Copilot has focused more on generative tasks such as summarising emails or drafting documents, Copilot Cowork takes on the execution stage of complex workflows.

Jared Spataro (재러드 스파타로), chief marketing officer for Microsoft’s AI at Work unit, explained in a blog post that Copilot Cowork handles long-running, multistep tasks.

Copilot Cowork will be offered through the Frontier program, which allows companies to test new features ahead of general availability.

When a user enters a goal, Copilot Cowork first develops a plan and then reasons through execution while moving across multiple Microsoft 365 apps and files. It is not fully unsupervised automation. Users can monitor progress and step in to adjust if the flow goes off track.

Microsoft said Copilot Cowork is based on the Work IQ framework, which accounts for security and governance compliance while learning the context of organisational data.

Microsoft aims to use Copilot Cowork to reduce inefficiencies caused by switching between apps. In the field, tasks such as monthly budget reviews require users to move frequently among apps such as Excel, Outlook, Teams and SharePoint to gather materials, coordinate with colleagues and complete reports. The company said Copilot Cowork bundles the process with an orchestrator, allowing tasks such as daily briefings and calendar management to continue without step-by-step prompts.

Copilot Cowork uses both OpenGPT models and Anthropic Claude. Depending on the user’s request, an appropriate model can be applied automatically, and users can also choose directly if they want.

Capital Group, an asset manager that was among the first to adopt Microsoft Copilot, said Barton Warner (바튼 워너), senior vice president for enterprise technology, said it “connects the steps, orchestrates the work, and processes the entire day-to-day workflow through to the end.”

Microsoft also revamped its Researcher agent alongside Copilot Cowork. The main change is the introduction of a “critique” layer that combines OpenAI GPT models and Anthropic Claude.

Under the structure, one model generates a draft and another model reviews accuracy and citation quality.

Microsoft said this approach raised its DRACO benchmark score, an industry metric for in-depth research quality, by 13.8 percent. The roles of the two models can be swapped, and users can also compare outputs from the two models through the model council feature.