lewis richards

How do you test Copilot against real-world business complexity without exposing customer data? Microsoft researchers created a platform that generates synthetic enterprise environments, complete with simulated employees, meetings, emails, documents, and industry-specific workflows. These environments let Microsoft evaluate Copilot at scale in realistic business scenarios while preserving customer data privacy. Read about the Industry Evaluation Data Generation Platform: msft.it/6011ahGkZ

markrussinovich.github.io/ve… #AI #OpenSource #MachineLearning #DecisionModels

How the Copilot Harness in Copilot Studio enables higher quality agents. microsoft.com/en-us/copilot/…

How do you solve the trust problem with AI-generated code? AI agents can generate code at a scale too great for humans to review line-by-line. That makes it difficult to ensure the code will do what the developer intended, and it leaves the final output vulnerable to bugs and security risks. Proof-Driven Development Kit turns plain-language developer intent into a specification in Verus, applying formal methods traditionally used for safety-critical systems. PDD Kit uses mathematical proofs to test agent-generated code against the spec where possible, and clearly indicates where the proofs can’t reach. The result is more reliable agentic code. Researchers @MatthaiP39719 and @alexlavaee break down Proof-Driven Development in the MSR blog: msft.it/6019af2V7

An agent shouldn't click when it can code. Same spreadsheet task: Qwen3.6-27B, clicking: 9:56 Microsoft's 4B Mage-VL-Cowork, coding: 0:39 It picks per step: 1. Layouts, canvases: click 2. Sheets, docs, slides: code 3. Checking its work: terminal huggingface.co/microsoft/Mag…

There isn’t one model that’s best at everything anymore. Choosing the right one for each task (without sacrificing quality or paying more than you need to) is a hard engineering problem. That’s what we’re working on with Auto in Copilot. Today, more than 90% of Copilot queries use Auto to dynamically route work across models and optimize for quality, speed, and cost. In this video, I share how it works and chooses the right model for the job...

We’re about 35 mins to showtime and the arena is filling UP! If you’re at #PPCC, get on over to the Garden Arena! There are some breakky burrito looking things, coffee and 9K of your closest friends filling up this room! I know @clamanna is going to kick off this morning with thanking the BEST community in the world. So much good stuff going LIVE today…there is a very specific announcement I’m waiting for 😅 Let’s go! 🚀

Jeff Teper's recap of the last three months is worth your time. It's incredible to see the pace of change and innovation, which means we MVPs have our work cut out for us to keep up.

For the past year we've focused on helping Copilot understand your workbook. The next step is helping it take on more of the work. You can now turn workbook data into an interactive canvas that stays connected as the numbers change, teach Copilot the processes you repeat with skills, and spend a lot less time getting data ready.

For the love of all that is holy, we don't need better models for AI adoption in business. We need to actually focus on solving problems for businesses. The models are already there. We're focusing on the wrong things. One-shotting video games with Astra doesn't move the needle.

I don’t say this often enough. The Microsoft MVP and AI people I follow here have changed how I work. Not in a vague “inspiration” way. In the practical way: I see a post, I try it, and I ship something better the same week. @FlowAltDelete @sharepointfocus @ShanesCows @rezadorrani @WonderLaura @DamoBird365 @MVPAward Thank you for putting the experiments, the failures, and the working patterns out in public. That kindness is doing more for working professionals than it ever gets credit for. You matter to me. You matter to a lot of us, many of whom would not have the careers we have without that generosity of spirit.

Excited to share VidMap! #ECCV2026 We bring the temporal structure of SLAM into global SfM for accurate 3D reconstruction of long, unconstrained videos. 🚀 Code: github.com/cvg/vidmap 📄 Paper: arxiv.org/abs/2607.27194 🎥 Video: youtu.be/gy4szd3q5Oo With @pesarlin & @mapo1

Agent Optimizer in Microsoft Foundry helps teams turn traces, evals, and production feedback into better agent configurations, with a loop built to generate candidates, score them, review changes, and ship the winner. See how the optimization loop works, when to use it, and how to try it today: msft.it/6015vYnWV