Speaker
Description
As AI agents move from short, inexpensive tasks toward realistic machine learning and scientific workflows, computational cost becomes an important part of the decision-making problem. Model training, evaluation, simulation, and high-fidelity validation can differ substantially in resource requirements, making exhaustive exploration impractical. SIDERIUS is a resource-aware multi-agent system for automated machine learning. It coordinates agents across scientific reasoning, model proposal, implementation, validation, tuning, and interpretation, while explicitly accounting for limited computational resources. Its design highlights the challenges that arise when agents operate over long and expensive ML workflows, and provides practical lessons from application to a real scientific machine learning problem. These lessons also motivate SciTra (Cost-Aware Science Trajectory), an ongoing next step toward cost-aware evaluation of long-horizon scientific agents.