Workshops Archive

Governed Agents: An Evidence-Based Architecture for the Computing Continuum


Workshop: 4th Annual Trillion Parameter Workshop: Building Open AI Infrastructure, Models, and Agentic Systems for Science

Authors: Vipin Chaudhary (Case Western Reserve University) and Raghu Machiraju and Dhabaleswar Panda (The Ohio State University)

Abstract: Agents can automate model selection, data curation, placement, retraining, and recovery across edge, cloud, and HPC resources, but this autonomy can turn uncertain model outputs into costly, privacy-sensitive, or irreversible actions. We define a governed agent as one whose proposed actions pass machine-checkable gates for context validity, decision uncertainty, policy compliance, execution feasibility, and auditability. Drawing on results from the NSF ICICLE AI Institute and related work, we present an architecture that integrates semantic typicality, runtime anomaly detection, neuro-symbolic verification, uncertainty-aware evaluation, lossless model compression, and modular agentic data curation. These mechanisms form a governance plane that mediates a sense-decide-act-learn loop across the computing continuum. Because current evidence remains component-level, we specify a falsifiable integration protocol, evaluation metrics, and failure tests to determine whether governed autonomy improves utility without obscuring risk or shifting excessive burden to human operators.


Back to 4th Annual Trillion Parameter Workshop: Building Open AI Infrastructure, Models, and Agentic Systems for Science Archive Listing

Back to Full Workshop Archive Listing