Transforming R&D with agentic AI: Introducing Microsoft Discovery

We have architected Microsoft Discovery to be highly extensible, enabling researchers to integrate the latest Microsoft innovations with their own models, tools, and datasets as well as a wide range of partner and open-source solutions. We are announcing a new enterprise agentic platform called Microsoft Discovery to accelerate research and development (R&D) at Microsoft Build 2025. Our goal is to bring the power of AI to scientists and engineers to transform the entire discovery process—from advanced knowledge reasoning and hypothesis formulation to experimental simulation and iterative learning. Microsoft Discovery enables researchers to collaborate with a team of specialized AI agents combined with a graph-based knowledge engine, to drive scientific outcomes with speed, scale, and accuracy. Get started by using Azure HPC and Azure AI Foundry infrastructure We have architected Microsoft Discovery to be highly extensible, enabling researchers to integrate the latest Microsoft innovations with their own models, tools, and datasets as well as a wide range of partner and open-source solutions. Built on top of Microsoft Azure, trust, compliance, transparency, and governance are key design principles of this enterprise-ready platform to enable responsible innovation, keeping the researcher in control. At Microsoft, our researchers have leveraged the advanced AI models and high-performance computing (HPC) simulation tools in Microsoft Discovery to discover a novel coolant prototype with promising properties for immersion cooling in datacenters in about 200 hours—a process that otherwise would have taken months, if not years. This rapid discovery lays the groundwork for future developments in safer and sustainable solutions across multiple industries and is a demonstration of how Microsoft Discovery can potentially transform R&D in any company. We are working with a notable set of Microsoft customers who are interested in co-innovating in diverse industries including chemistry and materials, silicon design, energy, manufacturing, and pharma. We are also working with a broad partner base that is building on top of the platform to drive this acceleration, and we couldn’t be more excited. The possibilities are endless as we realize the full potential of AI in R&D and we are just getting started! The agentic vision for science At Microsoft, we want to amplify the ingenuity of scientists to usher in a new era of accelerating discovery and expand the horizons of research. Doing so requires empowering R&D teams with transformative technologies that can drive meaningful business impact. However, R&D has very specific challenges compared to other domains: Scientific knowledge is vast, nuanced, and distributed. The discovery process is diverse and dynamic, involving multiple highly specialized methods and tasks, making it very hard to connect the dots across the different domains involved. R&D is iterative. There are rarely simple, clear-cut answers. Instead, scientific knowledge evolves through evidence, discourse, and refinement. This complexity demands a new paradigm—one that isn’t aimed at doing the same experiments faster, but rather fundamentally changing the paradigm of how we approach R&D. Imagine if every researcher could collaborate with a tireless team of intelligent, synergistic AI agents with the sole purpose of accelerated innovation. This is our vision for a new agentic R&D paradigm, embedding AI in every stage of the scientific method. In this new world, people and specialized AI agents will cooperatively refine knowledge and experimentation in real time in a continuous, iterative cycle of discovery—all while maintaining the control, transparency, and trust that enterprises and governmental institutions require. This requires a comprehensive platform where AI can capture both the scientific domain and the cognitive processes involved in managing scientific thought. To realize this vision, scientific AI agents must be able to: Reason over a complex and contextual graph connecting all knowledge sources. Specialize across distinct domains and tasks. Learn from results and adapt entire research plans accordingly.

Introducing Microsoft Discovery We are taking a big step toward realizing this vision with Microsoft Discovery, bringing agentic R&D to life by leveraging the latest innovations from Microsoft and the broader scientific ecosystem. Graph-based scientific co-reasoning ​ The advent of large language models (LLMs) hinted at this new era, offering capabilities to speed up certain scientific tasks, particularly for information retrieval and hypothesis generation. However, LLMs often lack the contextual understanding required to deeply reason over distributed, nuanced, and often contradictory scientific data.

Microsoft Discovery Get started today by using Azure HPC and Azure AI Foundry infrastructure. Learn more Microsoft Discovery is built on top of a powerful graph-based knowledge engine. Instead of merely retrieving facts, this engine builds graphs of nuanced relationships between proprietary data as well as external scientific research. This allows the platform to have a deep understanding of conflicting theories, diverse experimental results, and even underlying assumptions across disciplines. This contextual reasoning is also transparent. Rather than outputting monolithic answers, it keeps the expert in the loop with detailed source tracking and reasoning, providing the level of transparency in AI systems that builds trust, ensures accountability, and allows experts to validate and understand every step or make any adjustments as needed. Specialized discovery agents for conducting research Instead of siloed and static pipelines, Microsoft Discovery implements a continuous and iterative R&D cycle where researchers can guide and orchestrate a team of specialized AI agents that learn and adapt over time—not just for reasoning, but for conducting research itself. The definition of these specialized agents captures both domain knowledge and process logic, simply through natural language. R&D teams will be able to build a custom AI team aligned to their specific processes and knowledge, easily encoding these agents with their expertise and methodologies to ensure they can adapt and orchestrate as research progresses. This approach is far more flexible than hard-coding behaviors of today’s digital simulation tools, which often are highly specialized and lack streamlined integration with others, and it means that research teams no longer require computational expertise to drive impact. As an example, users can access and define various agents’ specialties, such as ‘molecular properties simulation specialist’ or ‘literature review specialist.’ They can even suggest which tools or models the agents should use or create, and how they should collaborate with others. This organic, bidirectional collaboration is a game-changer for managing R&D: agents are not only capable of working for the researchers, but with them in a manner that can truly amplify human ingenuity—seeing both the forest and the trees at once. At the center of this collaboration is Microsoft Copilot, acting as a scientific AI assistant that orchestrates these specialized agents based on the researcher’s prompts. Copilot is aware of all the tools, models, and knowledge bases in a customer’s catalog on the platform, can identify which agents to leverage, and can set up end-to-end workflows that cover the full discovery process by combining advanced AI and HPC simulations through the joint work of these agents. Extensible and enterprise-ready Microsoft Discovery is built on top of Azure infrastructure and services, leveraging by design the trust, compliance, and governance controls at the core of Microsoft’s secure cloud foundation. We believe in the power of an open ecosystem that leverages the strengths of Microsoft’s latest advancements in combination with other innovative solutions from customers and partners. Microsoft Discovery allows R&D teams to extend the platform’s catalog by bringing their toolkit of choice to cover their specific research needs in a comprehensive scientific bookshelf. This extensibility at the core of Microsoft Discovery simplifies the onboarding of their choice of computational tools, models, and knowledge bases—whether they are custom developments, open-source, or commercial solutions. As we bring to market new capabilities in reliable quantum computing and embodied AI, the platform will remain future-proofed with the best technologies available at Microsoft and across the industry. Real impact: Discovering a novel, non-PFAS coolant prototype Over the past months, we have made significant strides aiding computational scientists in their research and incorporating cutting-edge innovations from Microsoft Research. This has led to remarkable breakthroughs, such as discovering a novel solid-state electrolyte candidate that uses 70% less lithium in collaboration with the Department of Energy’s Pacific Northwest National Laboratory (PNNL) and enabling rapid computational simulations that accelerate scientific discoveries at Unilever. Microsoft Discovery is designed to bring these innovations to every scientist—not only those with deep computational expertise. One of the more exciting early use cases of Microsoft Discovery is unfolding at the Pacific Northwest National Laboratory, where scientists are using Microsoft Discovery’s advanced generative AI and HPC capabilities to further develop machine learning models that predict and optimize complex chemical separations—a critical process in nuclear science. These separations are essential for effectively isolating radioactive elements after the nuclear fission process, a notoriously time-sensitive and incredibly chemically complex task. In the future, the team aims to use these advancements to reduce the time scientists must spend in hazardous radioactive environments, while impro