Abstract
The rapid advancement of large language models has enabled powerful AI agents with reasoning, planning, and coding capabilities. However, agent behavior is shaped not only by the base model, but also by its harness—the surrounding system that mediates the model’s interaction with the environment. In this talk, we will present our recent explorations of agent harness. First, we will introduce Self Harness, a new paradigm in which an LLM agent acts as its own harness engineer, autonomously improving the harness that supports its operation. Second, we will present Harness of Harness, exploring how to scale autonomous software development from hours to days while enabling continuous improvement. Finally, we will briefly share our recent explorations of agent harness for theoretical computer science, investigating how harness-based approaches can enable new forms of AI-assisted research and discovery.
Time
Tuesday, Sept.1, 14:00-15:00
Speaker
Shuyue Hu is a researcher at Shanghai Artificial Intelligence Laboratory. Her research interests lie in agents, multi-agent systems, large language models, and game theory. She received her B.Sc. from South China University of Technology in 2015 and her Ph.D. from Chinese University of Hong Kong in 2019, under the supervision of Prof. Ho-fung Leung. Prior to joining the lab, she was a research fellow working with Prof. Harold Soh at the National University of Singapore and Prof. Georgios Piliouras at the Singapore University of Technology and Design. She has published over 50 papers in leading journals and conferences, including PNAS, National Science Review, Artificial Intelligence, AAAI, IJCAI, NeurIPS, ICML, and AAMAS. Her research has received the First Prize of the Natural Science Award from the Chinese Institute of Electronics, the Best Paper Award at DAI 2025, a Best Student Paper Award nomination at AAMAS 2026, and the Best System Demonstration Award at AAMAS 2026.
Yang Chen is a researcher at the Shanghai AI Lab. He obtained his Ph.D. in C from the University of Auckland and conducted postdoctoral research at both the University of Auckland and the University of New South Wales. He has been engaged in research in areas such as reinforcement learning, imitation learning, multi-agent systems, and game theory. His publications are on renowned international conferences in artificial intelligence such as NeurIPS, AAAI, IJCAI, AAMAS, and ACL. His research results have formed a relatively systematic "reward–decision reasoning" research framework, providing a new foundational perspective for the behavioral understanding, strategy prediction, and decision optimization of intelligent agent systems. He serves as the local chair of the top international conference on multi-agent systems, AAMAS 2024, and the chair of the AAAI Track. He also serves as a reviewer for top journals and conferences in the field of artificial intelligence, such as JMLR, ICML, NeurIPS, ICLR, AAAI, AAMAS, ACL, and IJCAI. He received the best demo award in AAMAS 2026.
Room
Room 104
