In the third installment of our Day in the Life series, we follow Microsoft researcher Ahmed Awadallah, a leader in agentic AI, as he navigates deep research, team leadership, and the evolving frontier of intelligent systems.

I’m currently focused on building agentic AI systems—models that can plan, execute, and adapt reliably and safely. These systems are designed to handle increasingly complex multi-step tasks, interact with their environments, and learn to improve over time, all while maintaining strong oversight, alignment, and transparency.

This work on AI agents really kicked off in 2023 with AutoGen, a project I helped lead. AutoGen is a widely adopted open-source framework for building AI agents and enabling multi-agent collaboration to solve tasks. Since its launch, AutoGen has graduated to Azure, where the team continues to evolve to provide enterprise-ready agentic AI capabilities on Azure.

Last year, Microsoft introduced the Phi family of small language models (SLMs), which redefined what’s possible with SLMs by achieving competitive performance with dramatically smaller footprints. Our current focus is on making these models agentic, starting with Phi-4-reasoning, which incorporates reinforcement learning and advanced synthetic data generation with multi-agent simulation to unlock stronger reasoning capabilities.

One misconception I often hear is that AI researchers spend most of their time inventing entirely new algorithms. In reality, much of the work happens elsewhere: curating data so it reflects the task accurately, designing experiments that test the right hypotheses, carefully evaluating results, and improving the infrastructure that makes training and testing more streamlined. 

Another thing I wish more people understood is that the kinds of tasks AI can and cannot do don’t always line up with human intuition. For example, people often assume that if an AI system can solve a complex math problem, then it should easily handle a much “simpler” common sense reasoning task. But the reality is that what feels easy for humans can be incredibly difficult for AI, and vice versa. This also relates to the misconception that if AI can do something a human can do—say, solve math problems or translate languages—it must be doing it in the same way humans do. The outputs can look impressively human-like, but the underlying processes are very different, which is why models can still make brittle mistakes in situations where a person wouldn’t.

I usually start from a problem or a goal rather than from a new method or an abstract curiosity. That might be a capability we don’t yet have or a limitation in existing methods. I try not to define the problem too literally at the beginning, since the right framing often evolves as we develop more understanding—but I also avoid moving the goal posts too often, because that can hinder progress.

Another important factor is whether the question allows for incremental progress. Ideally, I want to see a path where we can build and measure tangible steps toward the bigger goal, rather than waiting for a single breakthrough at the end.

Finally, I’m often drawn to problems that are both use-inspired and fundamental—a perspective shaped in large part by several mentors I had here at Microsoft, who emphasized the value of work that advances underlying science while also connecting back to real capabilities and applications.

Since early 2023, we’ve been enthusiastic adopters of Agentic AI and synthetic data generation, observing significant gains in improving SLMs by using large models as teachers to generate demonstrations and explanations for problem-solving (Orca). The Phi model family pioneered the use of synthetic data generation at scale for pre-training SLMs that rival much larger models in performance. More recently, with AgentInstruct, we’ve shown that multi-agent simulations can generate diverse, high-quality data by producing prompts, responses, and environments for validation. These methods have proven especially effective in unlocking superior reasoning capabilities in models like Phi-4-reasoning.

We are making a lot of progress in using these methods for even more complex, multi-step tasks—helping to train agents capable of reasoning and acting to solve more problems. One of the directions I am most excited about these days is the convergence of agent-based simulation, synthetic data generation, and reinforcement learning. The goal is to build scalable simulation environments for many tasks—where models not only create tasks and data, but also receive feedback on their actions. This represents a step toward self-play systems, enabling agents to learn from the outcomes of their actions within large-scale simulated environments.

I've been fortunate to have several mentors that shaped my approach to research, but if I were to choose only one, it would be Susan Dumais, whom I was lucky to work with over many years in different capacities: mentor, manager, and collaborator. Two lessons stand out in how working with her influenced my work: first is unwavering curiosity and commitment to a rigorous, thoughtful approach to research and experimentation. Second is her remarkable ability in selecting the right problems to work on, despite her often attributing it to luck.

On the other hand, one of the most meaningful aspects of leading my team has been how much I’ve learned from the people I work with. Each person brings a different perspective and strengths, and I’ve found myself adopting ideas and approaches from many of them. Sometimes it’s a new technical insight, sometimes it’s a creative way of framing a problem, and other times it’s the persistence and conviction when they believe strongly in a direction. Being surrounded by talented individuals working together toward an ambitious goal is one of the most rewarding aspects of my job.

We are already witnessing significant progress in both capabilities and adoption of AI agents, and I’m eager to see continued advancements in the near future—particularly in making them more reliable and trustworthy, so they can increasingly augment human capabilities in everyday life.

I also hope that we will make progress in making AI more accessible and affordable, reaching areas and communities that typically have limited access to technology—serving as tutors, healthcare advisors, etc.

Working on AI right now is one of the most exciting opportunities out there and we are lucky we have the chance to participate in shaping this technology. My advice is: