M365 Research

M365 Research advances enterprise-grade Agentic AI so Microsoft Copilot can become a trusted and skilled digital coworker embedded in the flow of enterprise work. We develop the foundational capabilities and systems that help agents understand people, goals, processes, and organizational context; execute complex work reliably; learn safely from validated experience; and operate with the performance and economics required at global scale.

Enterprise work spans people, documents, meetings, messages, applications, processes, permissions, and teams. Agents operating in this environment must assemble the right context, retain useful experience, reason across multiple steps, use tools safely, recognize failures, and remain under appropriate human control.

These capabilities also increase latency, infrastructure demand, and cost. M365 Research therefore treats agent quality, reliability, performance, and efficiency as one connected systems problem, spanning enterprise context, workflow traces, governed learning, realistic evaluation environments, validated repair, and efficient serving infrastructure.

Can Agentic AI understand and execute enterprise work well enough to become indispensable—and can it do so with the performance, efficiency, and cost required at global scale?

Agentic Innovation

Quality. Reliability. Trust.

Advances reasoning, planning, memory, skills, multi-agent collaboration, evaluation, and governed self-correction so agents can consistently deliver high-quality enterprise work outcomes.

• Enterprise context, and workflows

  High-quality agentic work depends on understanding the environment in which work occurs: people, teams, documents, meetings, messages, projects, business processes, permissions, and organizational norms. We develop governed and persistent context that enables agents to remember relevant information, learn recurring workflows and procedures, acquire and apply reusable skills, maintain freshness and provenance, and operate within appropriate boundaries.

  Memora | LegoMem | ACON

• Reasoning, evaluation, and realistic environments

  Agent progress depends on evidence that reflects actual enterprise outcomes. We develop task-grounded evaluation methods, realistic and privacy-preserving synthetic environments, and data-quality analysis that distinguishes failures caused by models, retrieval, tools, environments, data, or graders. This work helps identify low-signal training examples before compute is spent and supports shared quality bars for determining whether improvements will generalize to real product workloads.

• Reliable and self-repairing agents

  Enterprise agents must reason across complex workflows, recognize uncertainty, detect failures, identify their causes, and recover without silently producing poor outcomes. We develop diagnosis-to-repair loops that localize failures in execution traces, propose bounded changes to prompts, tools, or orchestration, and validate those changes against regression checks before deployment.

  AgentRx | AutoSaddler | Algorithmist

• Applied agentic systems

  We bring foundational capabilities into real productivity and engineering workflows where research can improve reliability, productivity, and efficiency. These settings include enterprise knowledge work, software modernization, workload knowledge, cloud operations, incident diagnosis, and account recovery.

  FLASH

Efficient AI

Performance. Cost. Scale.

Advances inference optimization, latency and cost efficiency, and hardware-software co-design so Agentic AI can operate responsively and economically across the Microsoft 365 ecosystem.

• Intelligent inference serving

  GenAI workloads differ substantially in context size, latency needs, quality requirements, safety requirements, caching potential, and execution patterns. We develop workload-aware serving techniques spanning traffic partitioning, endpoint tailoring, scheduling, batching, routing, load balancing, and configuration selection.

  GAUSS

• Workload-aware GPU and kernel optimization

  AI performance depends on the interaction among models, serving configurations, compute kernels, memory behavior, communication collectives, hardware topology, and workload shape. We co-design these layers around real Microsoft 365 workloads to improve throughput and useful work per GPU-second.

  LeanAttention | Capillary Kernels

• Scalable inference for complex and long-running workflows

  Reasoning models and long-running agents accumulate context, reuse memory, call tools, branch across steps, and operate over extended periods. We develop KV-cache lifecycle management, context-footprint reduction, asynchronous and throughput-oriented serving, adaptive compute, and workflow-level cost-and-quality measurement so complex agentic experiences remain responsive and economically sustainable.

  Prompt Compression | AutoSaddler | LegoMem

• Agent-assisted systems research and evaluation

  We use agents to help explore large configuration, scheduling, and kernel-design spaces while keeping researchers accountable for problem framing, validation, and interpretation. We also build reusable research knowledge and evaluation assets that help distinguish promising technical results from improvements that will not translate into customer or business outcomes.