Autonomous Discovery at Exascale Through Agentic AI and Multi-Agent Coordination

This invited talk presents work from the University of Chicago and Argonne National Laboratory group on scaling scientific productivity through AI-native science and multi-agent systems. The speaker argues that the complexity of the scientific enterprise is outgrowing human capacity, motivating a progression from the Parsl parallel workflow library and the funcX/Globus Compute function-as-a-service model to Academy, an open-source agentic middleware for deploying persistent, stateful agents across HPC systems, cloud resources, instruments, and edge devices while respecting data sovereignty. An agent is defined as a persistent process that observes, plans via a policy such as a language model, acts by executing tools or driving experiments, and learns. Application examples include a metal-organic framework discovery pipeline with generator, validator, and optimizer agents spanning the Aurora and Polaris supercomputers and Chameleon cloud; an automated plant-growth laboratory at Oak Ridge feeding computations to Frontier; a hypergraph-based system for predicting technology and policy outcomes; and a pandemic preparedness data exchange with the World Health Organization in Africa where agents must remain deployed within the region. Lessons drawn from an IPDPS paper include the need for federation, specialization of models and hardware, and dynamic distributed orchestration, since tool calling alone is insufficient for complex scientific workflows. In discussion, the speaker notes productivity gains remain qualitative, trust requires provenance and observability, vibe-coded components have introduced critical bugs requiring human oversight, and model costs relative to large computational campaigns remain modest.