Microsoft answers with Copilot Cowork
Microsoft’s reply came quickly. Copilot Cowork lets users hand off multi-step work that can move across Outlook, Teams, Excel, PowerPoint, and the rest of Microsoft 365, turning a request into a plan that continues in the background with checkpoints, clarifications, and approval moments along the way. Microsoft says Cowork is grounded in a user’s emails, meetings, messages, files, and data through Work IQ, runs within Microsoft 365’s existing security and governance boundaries, and is in Research Preview ahead of broader availability through its Frontier program later in March.
Why it matters: This is Microsoft acknowledging a basic truth about workplace AI: app-by-app assistance is no longer enough. The winning product has to move across the whole suite.
Read more
Google turns Gemini into a cross-workspace drafter
Google’s latest Workspace update pushes Gemini closer to being a true production assistant rather than a writing helper. Across Docs, Sheets, Slides, and Drive, Gemini can now pull from selected files, emails, and the web to draft documents, build spreadsheets, generate presentations, and fill in missing information from a single prompt. Google is also adding tools that let users match writing style and document format, making the system better at assembling a polished deliverable from the context that already exists around the user.
Why it matters: Google’s position here is unusually strong. If your work already lives in Gmail, Drive, Docs, and Sheets, Gemini does not need to win an integration war first. It is already inside the building.
Read more
Gemini Embedding 2 gives enterprise retrieval a multimodal upgrade
Beneath the flashier product launches, Google also shipped one of the week’s more consequential infrastructure updates. Gemini Embedding 2, now in public preview through the Gemini API and Vertex AI, maps text, images, video, audio, and documents into a single embedding space, supports mixed-modality inputs, and lets developers scale dimensions down from a default 3,072 to balance performance and storage costs. In plain English, it is a cleaner way to search and retrieve meaning across the messy, mixed-media reality of enterprise data.
Why it matters: Enterprise AI lives or dies on retrieval quality, context quality, and latency in the data layer. A stronger multimodal embedding model means fewer brittle workarounds, less pipeline sprawl, and better odds that an agent can actually find the right information when the task is no longer just text-in, text-out.
Read more
MCP is becoming the connective tissue for agentic software
Manufact, a startup building tooling around the Model Context Protocol, raised $6.3 million this week as MCP continues its rise from developer curiosity to serious infrastructure. MCP describes itself as an open-source standard for connecting AI applications to external systems, and Manufact is building SDKs and cloud tooling to help developers create MCP servers, apps, and agents that can work across major AI clients. The company’s broader wager is straightforward: as agents become real software users, every serious product will need an interface designed for them.
Why it matters: Every platform shift eventually needs a common connector. If MCP keeps spreading, the biggest opportunity may not be another model wrapper or chatbot skin, but the infrastructure that makes agent-to-software communication reliable, portable, and boring. And in enterprise technology, boring is usually when the market becomes real.
Read more
Anthropic pushes Claude deeper into the software development workflow
Anthropic has rolled out Code Review for Claude Code, giving teams a pull-request review tool that can use multiple agents to inspect code while leaving final approval to humans. The move builds on Anthropic’s broader push to make Claude part of the engineering workflow itself, including more automated security and review tasks that sit closer to real software operations than simple coding assistance.
Why it matters: Coding agents are moving beyond autocomplete into real engineering process. Once AI starts participating in code review and security review, model vendors are no longer just productivity tools. They become part of the software supply chain.
Read more