Microsoft just dropped what it’s calling a new era of doing work inside Word, Excel and Copilot chat: “vibe working.” If that sounds like vibe coding — where you sketch what you want in plain language and an AI scaffolds a working app — you’re on the right track. Microsoft’s new Agent Mode in Excel and Word turns a prompt into a multi-step plan and then executes it inside the document or spreadsheet, showing you each step. At the same time, a separate Office Agent inside Copilot chat (powered by Anthropic models) can build full Word docs and PowerPoint decks from a chat prompt. Both launches are being rolled out through Microsoft’s Frontier early-access program.

What it does

Think of Agent Mode like an assistant that does the fiddly, repeatable stuff while you act as director and verifier. Give it a prompt — “make a monthly sales report, show trends vs last quarter, highlight anomalies, and make three charts” — and Agent Mode will:

• break that brief into smaller tasks (plan),
• run each task (formulas, sheets, charts),
• show intermediate results and validation steps in a sidebar (so you can follow or correct them), and
• deliver a board-ready artifact in minutes.

Microsoft pitches it as more than “helpful” text suggestions; this is orchestration and execution — an automated macro that explains itself as it runs. That’s the vibe working pitch: hand routine, multi-step knowledge work off to an agent and keep the judgment and verification for humans.

Microsoft says Agent Mode in Excel leans on OpenAI’s latest reasoning models (Microsoft references GPT-5 in their materials) to “speak Excel” natively — deciding which formulas to use, generating new sheets, and making visuals. Microsoft published evaluation results using SpreadsheetBench, saying Agent Mode scored 57.2% accuracy on that benchmark (as of Sept. 29, 2025) — higher than several competitors Microsoft tested (Shortcut.ai, ChatGPT with .xlsx support, Claude Files Opus 4.1) but still behind human accuracy (71.3%). Meanwhile, the new Office Agent in Copilot chat is powered by Anthropic models and can assemble full PowerPoint decks with web research and live previews. Microsoft frames this as mixing and matching model families to get the best result for each task.

Where you’ll find it (availability and limits)

At launch, Agent Mode for Excel and Word is available on the web for Microsoft 365 Copilot customers and Microsoft 365 Personal/Family subscribers enrolled in the Frontier program; desktop support is promised “soon.” Office Agent in Copilot chat is also shipping in the Frontier program, but initially in the U.S. and using Anthropic models running on AWS for Copilot chat. Microsoft is also surfacing Agent Store and prebuilt agents inside the Copilot app so people can pick or customize workflows.

Why Microsoft is framing this as a pivotal shift

Two ideas are packed into the messaging:

1. Democratization of expertise. Excel has decades of arcane functions and workflows that only power users master. Agent Mode promises to make expert-level modeling accessible to non-experts by choosing formulas and validation steps for you — but, crucially, showing those steps so things remain auditable.
2. Agentic productivity. Rather than a one-off suggestion, these agents are meant to plan, execute, verify and iterate — a closer analogue to handing a junior analyst a brief and getting back drafts you can refine. Microsoft calls this “agentic” because the system takes initiative across multiple steps.

Where the friction will be (and what to watch)

This is a big step, but not a finished product. A few practical and policy issues are worth watching:

• Accuracy and auditability. 57.2% on SpreadsheetBench is a real improvement over many existing tools, but it’s still short of human accuracy. For businesses, that gap matters when models touch financials, forecasts, or compliance documents — places where an error can be costly. Microsoft stresses validation loops and visible sub-agent steps to make outputs verifiable, but teams will need internal checks.
• Data governance and security. Excel often holds sensitive corporate data. Making models act inside spreadsheets raises questions about data leaving tenant boundaries, where model inference runs, and how logs are stored. Microsoft’s messaging emphasizes enterprise controls and auditable artifacts, but IT teams will need to validate how the new flows fit their compliance posture.
• Model diversity and vendor mix. Microsoft is using OpenAI models inside Agent Mode and Anthropic models for Office Agent in Copilot chat — a deliberate multi-model strategy. That gives Microsoft flexibility (and negotiation leverage), but it also means different parts of Copilot may behave differently depending on which model family is chosen. Expect growing pains as teams learn where each model shines.
• Overreliance and skill atrophy. If agents reliably do junior consultant work, organizations must ask whether they’re losing human expertise over time. That risk is manageable if agents are used to augment and scale people — but dangerous if verification becomes perfunctory. Industry-wide, we’ll likely see new best practices for “agent supervision.” (This is more of an expectation than a Microsoft claim.)

Real examples Microsoft highlights

The company’s marketing shows straightforward use cases: run a full analysis on a sales dataset and get insights and charts; turn customer feedback into a trends summary; produce a monthly report comparing this month to the last; or have Copilot draft a slide deck with a research assistant that pulls web sources and previews slides live in the chat. In Word, Agent Mode shifts writing into an iterative conversation instead of a one-pass output — Copilot asks clarifying questions and suggests structure and phrasing as you go.

Competition and the market context

Microsoft isn’t the only company pushing agents into productivity apps, but the combination of Office’s entrenched user base and Microsoft’s multi-model strategy gives it a leg up. Competitors from Google to smaller startups are racing to bake agents into workflows, but Microsoft’s bet is that being able to create auditable Excel models, familiar Word docs and previewed PowerPoint slides inside the same Copilot ecosystem will be sticky for teams.

Final note

“Vibe working” is a tidy slogan for a genuinely consequential product shift: AI that plans, executes, and explains multi-step office work inside the apps people already use. It won’t replace humans for high-stakes judgment, but if Microsoft’s auditable agents can reliably handle a chunk of repeatable work, the way teams allocate time could change quickly. The accuracy gap to humans — and the governance around data and model choice — will shape how fast organizations move. For now, Microsoft is inviting users in to try it, temper enthusiasm with verification, and see whether vibe working truly becomes the new normal.