The agentic paradigm in AI gives rise to a new frontier of intelligence where the knowledge is not stored in a single model, but rather collectively among a population of collaborating agents and humans. When used in areas requiring novel discoveries, this framework encourages exploration and shares much in common with how networks of scientists collaborate to make new discoveries. Proper integration of such systems into lab-in-the-loop platforms promise to enable more effective experimentation, more rapid discovery, and new forms of human-AI collaboration.

Building multi-agent orchestration into every element of the lab-in-the-loop workflow. This means agentic harness designed specifically to enable tight collaboration between researchers and agents when performing experiments, collecting data, assembling datasets, and analyzing results; and multi-agent systems that collaboratively explore large search spaces efficiently, behaving like a community of scientists who communicate with their human counterparts effectively throughout the discovery process. That same multi-agent system iteratively refines and improves the models trained on the data the lab in the loop collects. Every element feeds back on the others, enabling the platform to improve autonomously.

Joseph most recently worked as a research engineer for Yoshua Bengio’s group at Mila - Quebec Artificial Intelligence Institute, working on open source library development, antibody loop optimization, and efficient distributed training frameworks for generative flow networks. He is a PhD student researching methods for autonomous discovery at the University of Montreal, and holds dual master’s degrees in Neuroscience and Computer Science. He’s previously worked as a researcher for Mila’s Applied Machine Learning Research Team, Deep Genomics, and the Center for Addiction and Mental Health, as well as interned at Imagia Medical Imaging team and Google’s search ad team. He’s also served as a technical mentor for Creative Destruction Lab Montreal and area chair for NeurIPS.