What's next in AI? Field notes from Microsoft Research for 2026

AI as a lab assistant helping to accelerate scientific discovery

Autonomous agents will transform digital economies

As AI agents evolve from isolated tools to active participants in our digital ecosystems, we stand at the threshold of a new economic era-one where autonomous agents collaborate, negotiate, and transact on behalf of people and organizations. These agentic ecosystems promise to reorganize digital marketplaces, reduce friction, and broaden access to opportunity. Realizing this vision requires rethinking the architecture of systems, platforms, and protocols that underpin digital markets with an agent-native lens.

• Traditional markets depend on human attention and platform intermediation, creating bottlenecks and inefficiencies. Agentic marketplaces introduce direct agent-to-agent negotiation and exchange, shifting incentives toward meaningful outcomes and enabling scalable, value-based economies. Yet this transformation introduces new complexities: agents must coordinate under information asymmetry, resist manipulation, and collaborate effectively with each other and with humans as core economic actors.

  We investigate these futures through simulation frameworks such as Magentic Marketplace, which models two-sided agentic markets and stress-tests agents on dimensions of trust, security, and collaboration. Early experiments show promise but also reveal challenges like systematic biases, adversarial tactics, and coordination failures.

  In the coming year, we'll be focused on developing behavioral protocols, collaborative models, and oversight mechanisms to ensure fairness and resilience in agent-driven economies.

AI meets biology to decode life's language

For decades, computational biology operated within narrow lanes, predicting protein structures or analyzing gene expression, while medicine largely treated patients as averages. What's new is the rise of generative AI models that treat biology as a language, enabling systems to design new proteins and predict cellular behaviors that can lead to personalized therapies.

• At Microsoft Research, we see this as an opportunity to move beyond static models toward architectures that integrate across biological modalities via generative reasoning. Our public work spans massive datasets like the Dayhoff Atlas and models such as EvoDiff, which learn from billions of protein sequences to create biomolecules never seen in nature, as well as the joint research initiative Project Ex Vivo(opens in new tab), which bridges computation and experimentation to define and target cell states in cancer. These advances could accelerate drug discovery and bring precision medicine closer to reality.

  The implications are enormous, but the progress isn't without complexity. Data quality remains critical; biology is complex; and real-world translation is everything. Even with these challenges we're driving innovation forward by revealing what works and where to focus next.

Future AI infrastructure enabling the next 1,000x

In 2026, two forces will redefine AI infrastructure. First, AI-driven system intelligence(opens in new tab) will unlock a step change in efficiency and velocity through automated tooling for developing, deploying, and optimizing models, all co-designed with the underlying hardware. Early signals are already here, with tooling accelerating adoption of models optimized for the edge(opens in new tab).

• Second, hardware disaggregation will break monolithic designs as specialized compute and bandwidth-optimized chips work in concert across workflows. Innovations across the stack, from compilers to optical interconnects, will enable this shift. Breakthroughs in optical communication, such as wide-and-slow interconnects based on microLEDs, will also ease cooling and packaging constraints in ultra-dense racks and unlock entirely new layouts.

  Beyond silicon, the horizon is even bolder: light-based chips(opens in new tab), new memory technologies, and robotics-enabled datacenter designs promise an era where AI infrastructure is faster, more sustainable, and fundamentally different. These longer-term advances are not optional. They are essential to meet surging demand, a Jevons paradox for AI that will continue to drive infrastructure investment.

Scaling AI at the speed of light

AI scaling is entering a new chapter. The challenge involves moving data quickly across GPUs and between GPUs and memory without burning too much energy. But new solutions are emerging. Throughout 2025, we've seen advances in low-power, high-bandwidth optical interconnects, from our own microLED work to developments across the GPU(opens in new tab) and networking(opens in new tab) ecosystem. As these technologies mature, I expect 2026 to be a pivotal year in moving them from R&D to early deployments, with wider adoption by decade's end.

• The impact could be transformative. Beyond addressing today's memory and network bottlenecks, these technologies will unlock new AI cluster designs. We could move away from power-hungry GPU racks toward smaller compute modules paired with shared memory pools, all connected by a fast optical fabric. This would enable a "disaggregated" AI datacenter where compute and memory resources can be pooled, composed, and reconfigured based on workload needs.

  Such flexibility could also transform the AI models themselves. Breakthroughs in interconnect technology, from the early internet to cloud networking, have led to new application paradigms. A future hyper-connected datacenter could similarly open the door to new classes of AI models that are smarter and more environmentally sustainable, sparking research directions we cannot yet imagine.

AI that amplifies human agency through inclusive innovation

AI's next frontier is not just smarter algorithms and models but the systems that amplify human agency in high-stakes domains like education, agriculture, and healthcare. The challenge: design AI-native workflows that serve a teacher or student in rural India, a farmer in Kenya, or a frontline health worker in Brazil. The answer lies in building AI that closes opportunity gaps and creates pathways for empowerment rather than reinforcing existing divides.

• We envision AI as a collaborator, not a mere tool. Imagine learning assistants that understand current learning levels and styles, the local context, curricula, and languages, and use this information to navigate the best learning path. Consider an agricultural advisor that integrates satellite imagery and market insights with community knowledge, guiding harvesting decisions and sustainable practices, all within a seamless interface that a farmer can easily use.

  Achieving this requires breakthroughs that go beyond accuracy to measure societal impact and interdisciplinary research that creates systems that are inclusive by design.

  We're building a future where AI doesn't replace human judgment but augments it, enabling globally diverse communities to thrive. By starting with the hardest use cases, we aim to set a new standard: AI that is equitable, collaborative, and transformative.

From reasoning to simulating and mentalizing

Every groundbreaking scientific revolution that transforms our lives begins with asking profound questions. Inquiries about the fundamental components of the universe, the measurement of information, and the origins of life have led to the discovery of three essential elements that define our physical and digital worlds: the atom, the bit, and the gene, respectively.

• One of the most intriguing questions in AI is the origin of (machine) intelligence. While a universally accepted answer remains elusive, recent advances in AI present new opportunities to explore this mystery. The AI community is now shifting its research focus from merely encoding world knowledge through large language models (LLMs) to developing reasoning abilities by enabling AI models to interact with their environments.

  We are pushing the boundaries of AI by developing models that go beyond logical reasoning to include simulation and social reasoning. These advanced models not only use world knowledge and trial-and-error learning but also possess the ability to simulate the external environment through an internal model, known as a world model. Additionally, they can understand the human mental state, a capability referred to as mentalizing. This enables AI agents to interact effectively with humans, infer user intentions, and collaborate more effectively.

Audience-shaped stories in real time

Looking ahead, AI is poised to transform entertainment into deeply interactive and personalized experiences. We'll move beyond static content toward dynamic worlds that adapt in real time to player choices, emotional context, and even social interactions. Generative models will empower creators to co-design with AI, blending narrative, art, and technology in ways that feel truly alive. Imagine stories that evolve with audience feedback, characters that learn and grow over time, and immersive environments that respond to human creativity and curiosity.

• What excites me most is the potential for AI-supported entertainment to create entirely new ways for us to connect. These experiences won't just be about playing a game or watching a film, they'll become shared spaces where creativity and collaboration flourish. Friends across the globe could co-create evolving narratives, communities could shape worlds together, and audiences could influence stories in real time. This convergence of AI and entertainment will redefine how we connect, play, and create - not as passive consumers, but as active participants in living, breathing worlds.

  I can't wait to see what becomes possible when technology and imagination combine to make entertainment more human, more social, and more inspiring than ever before.

AI as a partner to stabilize workflows and accelerate innovation

In 2026, AI stops acting like a tool you query and starts behaving like a collaborator you build momentum with. The breakthrough that enables this shift will be memory. Agentic systems will hold context across months, track evolving goals, surface forgotten assumptions, and help teams stay oriented in the messy middle of innovation. This reduces the cognitive tax of complex work and makes iteration feel continuous rather than fragmented. With AI as a stabilizing partner, teams will be able to move faster, and organizations will be able to preserve the intent behind their best discoveries, not just the artifacts.

Adaptive and collaborative robotics

Physical AI, where agentic AI meets physical systems, is poised to redefine robotics in the same way generative AI has transformed language and vision. For decades, robots have excelled in structured environments like assembly lines and warehouses, where tasks are predictable and tightly scripted. What's new is the emergence of AI "vision-language-action" models for physical systems that will soon perceive, reason, and act in dynamic environments alongside humans. These models translate natural language prompts into physical behaviors, enabling robots to adapt their actions by generalizing experiences across familiar scenarios with subtle variations rather than failing the first time they are presented with unexpected situations.

• At Microsoft Research, we see this as a foundational shift: building on classical control and reinforcement learning toward multimodal generative architectures that treat action as a first-class modality. Our work reflects this trajectory, combining spatial intelligence, tactile sensing, and generative reasoning to unlock robotic manipulation and collaborative teaming between humans and physical systems. The implications are profound: general-purpose robots that learn across tasks, interoperate in diverse environments, and serve as true partners in accelerating operations from datacenters to wet labs. This isn't some distant possibility; it's happening now as the next step in AI's evolution, as physical intelligence becomes the frontier for innovation and impact.

The many ways AI will reshape communication

Today, AI systems can generate text, images, and videos at an unprecedented scale, making information more accessible than ever. Yet the human capacity to interpret and internalize this growing stream of information is limited, creating a widening gap between content generation and comprehension.

• The next evolution of AI will focus on enhancing communication itself. Rather than packaging ideas into static artifacts such as documents or slides, AI will help sustain context, clarify ambiguity, and refine expression through continuous interaction. Communication will increasingly unfold as an iterative process, more closely aligned with how human thinking naturally progresses.

  We refer to this emerging paradigm as agentic media. Instead of serving merely as passive containers of information, media channels can become active participants in communication, forming dynamic environments that track how ideas evolve, reveal underlying reasoning, and adapt their structure to different audiences and stages of workflow.

  In 2026 and beyond, AI will not only increase what we can produce but also help ensure that what is created can be truly understood.

Improved context management helps AI act, not just respond

The next wave of AI will go far beyond question-answering and content generation. Agents will perform increasingly complex, long-running tasks that rely on external tools and real-time data. This shift introduces a critical challenge: preserving coherence and direction over extended sequences of actions. Agents will generate and consume far more information than a single prompt can hold, so "context engineering" will become essential, dynamically curating and structuring instructions, tools, and memories to keep the system on track. Improved state management will help today's promising agents become tomorrow's dependable collaborators.

AI empowering billions by turning vision into reality

AI has seen breathtaking growth in capabilities and adoption. It is important to also take a global perspective, and we're taking a three-pronged approach.

The language of the patient as a pathway to precision health

Today, each of us enjoys healthcare far superior to what any king or queen could have imagined centuries ago. That raises a profound question: can we enable every person on the planet to access care better than the most resourced individual today? With the rise of generative AI, we are beginning to knock on that door.

• Traditionally, medical discovery advances one clinical trial at a time. A cancer trial may take years and cost over $100 million, yet yield only a few thousand examples. By contrast, every day we collect billions of data points documenting patient journeys, from diagnosis to treatment to outcome. The challenge is that such journeys are difficult to interpret. A cancer patient's story spans radiology, digital pathology, and multiomics, each modality offering precious insights that sound like noise unless we understand the multimodal language behind them.

  By harnessing generative AI to master "the language of patients," we hope to one day develop "virtual patients" for precision health, digital twins that simulate disease progression and counterfactual treatment responses. This is thrilling, but AI is no magic wand. Real breakthroughs come when we reimagine the entire ecosystem and workflows, and embrace a future where multimodal intelligence transforms care at scale.

Psychological well-being as a core element of AI design and governance

As AI systems permeate everyday lives, their influence on psychological well-being is no longer theoretical. AI is mediating decisions, relationships, and narratives across personal, work, education, and civic spheres, and shaping how people think, feel, behave, relate, and understand themselves.

• The bold agenda for 2026 is to embed psychological well-being as a core AI design principle, not an afterthought. This means anticipating and mitigating psychosocial risks of mental health exacerbation, dependency, social fragmentation, and erosion of human dignity, while ensuring AI cultivates critical thinking, self-reflection, and healthy human connections. Psychological stewardship must be multidimensional, safeguarding individual resilience, fostering trust and empathy in interpersonal dynamics, and reinforcing societal cohesion. Achieving this is not just a technical feat. It demands partnerships across engineering, research, policy, and advocacy to define standards, tools, and accountability frameworks that directly translate research insights into actionable policy and implementation. The opportunity is clear: lead the industry in defining norms for psychological flourishing with AI, shaping a future where technology strengthens rather than destabilizes the human mind.

AI as a trusted companion in society

In 2026, we anticipate a major shift in how AI interacts with people, not as a tool that executes tasks, but as a trusted companion that collaborates, reasons, and grows alongside us. Instead of focusing only on accuracy or efficiency, AI systems will behave in ways consistent with human expectations for cooperation, respect, and social intuition.

• Our Value Compass project demonstrates how the structure of values can be empirically explored across cultures, offering early insights into how AI might understand and navigate the diversity of human preferences and social expectations. Meanwhile, our research on psychometrics-inspired evaluation suggests that we will soon measure AI not only for what it knows, but for how it interacts, adapts, and builds long-term trust.

  Looking ahead, we predict AI companions capable of maintaining shared histories and evolving relational styles, supporting individuals in work, creativity, and everyday decision-making. They will explain trade-offs, anticipate needs, and negotiate goals in ways that feel natural and respectful. As described in our Societal AI vision, realizing this future requires interdisciplinary collaboration in fields like psychology to understand human cognition and emotion, sociology to reason about communities, ethics and philosophy to guide judgment, and computer science to build robust systems.

Multimodal foundation models and agentic systems for patient care

The next phase of AI for healthcare will be defined by multimodal and agentic capabilities. Multimodal foundation models are beginning to integrate text (such as clinical notes), medical images, physiological signals, and genomics into a unified representation space for enhanced patient care. This enables systems to move beyond narrow clinical diagnoses toward broader, context-rich clinical reasoning.

• Clinical agentic AI is also evolving from passive copilots to workflow-aware assistants. Instead of only drafting summaries, future agentic systems may support triage, diagnostics, treatment planning, and coordinated follow-up. However, agentic behavior introduces new risks that must be mitigated through clinician-in-the-loop validation.

  Responsible translation will combine foundation model adaptation and task-specific fine-tuning with rigorous clinical evaluation. Agentic workflows will incorporate stepwise reasoning, uncertainty estimation, and reinforcement learning from clinician feedback to ensure safe integration into clinical workflows.

  True breakthroughs require deep collaboration to address real-world needs that go beyond benchmark performance. This aligns with our focus on AI for Industry by co-designing practical solutions and scalable platforms that meet various constraints while enabling meaningful transformation across healthcare settings.

System intelligence as the next leap

The future of computing isn't just faster-it's smarter. System Intelligence(opens in new tab) moves beyond AI's ability to generate code toward designing, optimizing, and governing entire systems. Instead of static architectures, imagine infrastructures that adapt and evolve toward high-level goals.

• To realize this vision, we must define and measure(opens in new tab) what intelligence in systems truly means, capturing how AI reasons about architectures, trade-offs, and correctness. This marks the shift from automation to autonomy.

  This transformation will redefine trust and scale. As complexity outpaces human capacity, autonomous systems will unlock breakthroughs in reliability, security, and performance-embedding intelligence into the very fabric of computing.