Microsoft Discovery puts AI agent teams to work in scientific R&D

Microsoft is turning AI agents into digital research assistants.

At Microsoft Build 2025, the tech giant introduced Microsoft Discovery, a new enterprise agentic platform designed to overhaul corporate and scientific R&D. Instead of researchers manually stitching together papers, scripts, and simulation tools, the platform lets them direct a team of specialized AI agents working alongside a graph-based knowledge engine.

A collaborative team of scientific agents

Traditional LLMs often struggle with frontier science because real research is messy, nuanced, and rife with contradictory experimental data. To tackle this, Microsoft Discovery relies on a graph-based knowledge engine that maps complex relationships across both proprietary laboratory data and external scientific literature. Every reasoning step and source is tracked, allowing researchers to inspect exactly how the AI arrived at a conclusion.

Sitting on top of that graph is a collaborative network of natural-language-configured agents. A lab can set up dedicated agents for tasks like molecular property simulation or literature reviews. Microsoft Copilot acts as the central conductor, identifying which agents and computational tools to deploy based on researcher prompts and orchestrating end-to-end simulation pipelines.

To give it extra computational punch, Microsoft is also integrating Nvidia's specialized microservices, including the ALCHEMI NIM for chemical simulations and BioNeMo NIM for drug discovery workflows. Under the hood, these workloads can run on Azure's new ND GB200 v6 virtual machines, powered by Nvidia Blackwell racks that Microsoft claims deliver up to 35x higher inference throughput than previous H100 systems.

Cutting years of chemistry down to 200 hours

Microsoft didn't just announce the platform—it brought a concrete proof point. Researchers inside Microsoft used Discovery's AI models and HPC simulation tools to identify a promising, non-PFAS coolant prototype for datacenter immersion cooling in about 200 hours, a search that normally takes months or years.

Early enterprise and government adoption is already taking shape. Scientists at the Pacific Northwest National Laboratory (PNNL) are using the platform to optimize machine learning models for chemical separations of radioactive elements after nuclear fission, aiming to cut down human exposure to hazardous environments.

Why it matters

We think Microsoft's approach here is very smart: rather than trying to build a single, opaque "frontier science" model that claims to solve chemistry, it built an orchestration layer. It bridges the stubborn gap between wet-lab scientists and high-performance computing clusters.

For enterprise R&D teams in materials, energy, and biotech, this drastically lowers the bar to entry for advanced computational screening. Researchers don't need to be code wizards to run complex HPC jobs; they just need to prompt and oversee an agent squad that knows their catalog.

That said, the stakes in physical science are unforgiving. While source attribution helps, relying on agentic workflows to interpret conflicting literature carries the lingering risk of hallucinated assumptions that could burn expensive lab validation budgets. But if platforms like Discovery can routinely compress months of candidate screening into a single week of compute, AI-driven labs are about to lap their traditional peers.