Agentic AI marks a shift from passive systems to active collaborators—AI agents that engage in sustained reasoning, understand complex multimodal environments, and interact naturally with humans over extended time and contexts. Our vision is to build intelligent systems that participate meaningfully in knowledge discovery, content creation, communication, and decision-making.
If you are fascinated by the future where AI evolves from tools into true partners, we warmly invite you to join the Microsoft Research Asia StarTrack Scholars Program. Here, you will explore the core capabilities that define tomorrow’s agents: sustained reasoning, multimodal perception, fluent human-AI dialogue, and open-world collaboration. Applications are now open for the 2025 program. For more details and to submit your registration, visit our official website: Microsoft Research Asia StarTrack Scholars Program – Microsoft Research.
Research Directions: Three Interrelated Pillars
We organize our understanding of Agentic AI into three interrelated categories, each reflecting a different facet of the ecosystem these systems must inhabit:
1) Foundations & Frameworks
Agentic AI requires new computational foundations to operate effectively across time, context, and modalities. We seek to develop compact and grounded multimodal representations that allow systems to perceive and reason over complex visual, auditory, and sensor-rich environments.
To support long-term engagement and contextual reasoning, we explore advances in semantic memory and process memory—compression mechanisms that allow agents to retain knowledge across interactions and reason over long horizons. Retrieval-augmented pipelines further enrich reasoning with dynamic access to knowledge bases and structured memory.
Agents should also coordinate and plan over extended workflows. We encourage research on multi-agent collaboration, process-aligned action spaces, and models that align with the semantics of user-driven tasks. Contributions in this category might include novel architectures, datasets, training methods, or theoretical insights that strengthen the core reasoning and planning capabilities of agentic systems.
2) User Experiences and Human-Agent Interfaces
Human–agent interaction must evolve to meet the demands of fluid, multimodal collaboration. We envision interfaces that are generative and dynamic, constructed to suit user intent and task context. Proposals should consider how agents generate or adapt interfaces in real time, across devices, modalities, and immersive environments.
Agents should be capable of interactive visualization, using media to communicate internal states, uncertainties, and possible outcomes. Interaction should be audience- and context-adaptive, with personalized behaviors that respect different usage settings and preserve user privacy.
We also encourage work on process-aware collaboration, where agents sustain memory across sessions and adapt over time. Topics such as communicative effectiveness, trust, interpretability, and longitudinal studies are central. Researchers should also consider how interface design, memory, and communication strategies can support natural, explainable, and long-term agentic interaction.
3) Applications & Societal Impact
We see Agentic AI as a transformative force across domains like science, healthcare, education, and enterprise decision-making. These applications require agents to perform deep research—gathering evidence, testing hypotheses, and constructing trustworthy narratives with transparency and provenance.
In the media domain, agentic AI systems could enable advanced content generation, curation, and personalization by understanding and synthesizing information from multiple modalities such as text, audio, and video. These agents support content creators by automating research, summarization, and fact-checking, or by generating interactive and adaptive media experiences tailored to audience preferences. Additionally, agents could facilitate media analysis at scale—detecting trends, ensuring provenance, and helping organizations manage and distribute content more effectively.
Finally, as these systems scale, we must ensure they operate responsibly. We encourage proposals on provenance tracking, watermarking, privacy-preserving design, and energy-efficient deployment. We also welcome organizational and societal studies that examine how agentic systems are adopted, trusted, and evaluated in the wild.
An Introduction to InfoAgent: Advancing Autonomous Information-Seeking Agents
Building on this strategic vision, Microsoft Research Asia has been actively turning the Agentic AI framework into reality through concrete, large-scale collaborative projects. A flagship example from 2025 is InfoAgent — a deep research agent designed to autonomously plan, search, and reason across the web to answer complex information-seeking queries.
InfoAgent was developed through close collaboration among multiple Microsoft teams and contributors from leading universities such as Southeast University and Brown University. With contributions from more than ten people, including Microsoft FTEs and interns, the project exemplifies the spirit of collaborative research in agentic AI, integrating expertise across LLM reasoning, reinforcement learning, data synthesis, and large-scale system design. We invite academic partners to join us in advancing next-generation agentic AI, deep research, and tool-augmented reasoning.
To address the key bottlenecks in building deep research agents, known as data quality and interactive environment, we focused on two major components that lay the groundwork for effective model training:
1)A data synthesis pipeline that constructs complex, multi-entity search questions. By building entity trees from Wikipedia and fuzzifying facts, the pipeline produces queries that demand long-horizon reasoning and tool use.
2)A self-hosted web search and browsing infrastructure that supports scalable, high-concurrency access to real-world information. This setup ensures transparency, removes dependency on commercial APIs, and enables reproducible reinforcement learning training.
InfoAgent was post-trained from Qwen3-14B through a two-stage process:
1.Supervised Finetuning (SFT): to instill multi-step reasoning and search behavior.
2.Reinforcement Learning (RL): to refine tool use and enhance decision-making efficiency.
Despite its modest 14B parameters, InfoAgent outperforms much larger open-source models such as WebSailor-72B and DeepDive-32B on multiple benchmarks. It achieves 15.3% accuracy on BrowseComp, 29.2% on BrowseComp-ZH, and 40.4% on Xbench-DS, establishing a new state-of-the-art among open models under 15B parameters. Remarkably, even trained only on English data, InfoAgent demonstrates strong cross-lingual generalization to Chinese benchmarks.
Beyond its technical achievements, InfoAgent lays a foundation for open, reproducible deep research environments. It shows how data synthesis, reasoning, and system engineering can co-evolve to push the frontier of AI autonomy.
Dr. Zhenghao Chen: StarTrack Scholar Story
While InfoAgent demonstrates the power of large-scale collaborative engineering in autonomous deep research, the StarTrack Scholars Program equally emphasizes nurturing exceptional individual researchers who bring fresh ideas and drive targeted breakthroughs within the Agentic AI ecosystem.
Dr. Zhenghao Chen, Assistant Professor in the School of Information and Physical Sciences and Centre for Applied and Responsible AI at the University of Newcastle, Australia, joined Microsoft Research Asia in 2025 as a StarTrack Scholar. Below is his personal reflection on this transformative collaboration:
My collaboration with the research group at Microsoft Research Asia (MSRA) marks a significant academic and personal milestone. It has strengthened my capabilities across multiple research areas, particularly in agentic AI for next-generation human–media interaction, positioning me to conduct world-leading research and to translate these advances back to Australia for tangible national impact.
Working with Dr Yan Lu’s group, I broadened my research through exposure to diverse, cutting-edge projects and their deployment in complex real-world enterprise settings. In particular, I was fortunate to contribute to a large agentic AI project. With the support of this esteemed team, I led the development of an agentic generative compression method that addresses semantic–memory bottlenecks in agentic AI and enables more efficient, data-in-the-loop interaction. This work represents a first step towards enabling AI agents to transform excessive raw multimodal data into compact representations that can be effectively communicated across agents, communication media and humans, while remaining perceptually faithful to people and reliably interpretable by AI models, supporting deployment across a wide range of devices.
Having the opportunity to contribute to agentic AI research with far-reaching implications for multimedia communication and interaction across large-scale platforms, I will bring this experience back to Australia to guide my translation agenda in line with Australian national priorities. Specifically, I am committed to using advanced agentic AI to develop an intelligent, immersive and interactive (i3) system, and to deploy this i3 system for: (1) Indigenous Knowledge – building National Aboriginal and Torres Strait Islander knowledge systems that enable the real-time preservation, sharing and interaction of cultural heritage and stories across platforms, under respectful data governance; (2) Regional Education – delivering efficient, offline-first interactive learning resources that support self-paced learning and locally relevant content, helping to mitigate educator shortages and connectivity or infrastructure constraints in regional communities; and (3) Environmental Support – enabling more efficient visualization, analysis and interaction with virtual scenarios of dynamic environments, such as flooding and bushfire, allowing scientists to explore, forecast and plan in highly realistic yet safe conditions.
Overall, my visit to MSRA advanced my scholarship and, more importantly, forged relationships that will shape my work for years to come. It was an honor to work with a world-leading research team. Guided by our team’s strategic frameworks and hands-on guidance, I extended my research design and broadened its impact. With generous support from colleagues, I saw MSRA’s inclusive values through open collaboration and thoughtful day-to-day care. The journey is not only an important academic chapter, but also a foundation for lifelong perspective, community, and partnership. Looking ahead, I will continue to collaborate with Dr Yan Lu’s group on high-impact, world-leading publications and will actively explore a formal Australia–MSRA partnership, for example through the Australian Research Council (ARC) Linkage Projects scheme, to accelerate research translation for both global and national impacts.
Microsoft Research Asia StarTrack Scholars advocates an open attitude, encouraging dialogue and joint experimentation with researchers from various disciplines to discover viable solutions. Now visit our official website to know more: Microsoft Research Asia StarTrack Scholars Program – Microsoft Research
- Yan Lu, Partner Research Manager, Microsoft Research Asia
- Chong Luo, Sr. Principal Research Manager, Microsoft Research Asia
- Jiahao Li, Principal Researcher, Microsoft Research Asia
- Yun Wang, Senior Researcher, Microsoft Research Asia
- Xun Guo, Principal Researcher, Microsoft Research Asia
- Kai Qiu, Senior Researcher, Microsoft Research Asia
- Qi Dai, Principal Researcher, Microsoft Research Asia
- Bei Liu, Senior Researcher, Microsoft Research Asia
If you have any questions, please email Ms. Yanxuan Wu, program manager of the Microsoft Research Asia StarTrack Scholars Program, at v-yanxuanwu@microsoft.com
Artificial intelligence