Workshop: AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC
Authors: Valesca Moura (Fluminense Federal University), Gabriele Padovani (University of Trento), Daniel de Oliveira (Fluminense Federal University), and Rafael Ferreira da Silva and Renan Souza (Oak Ridge National Laboratory)
Abstract: AI agents governed by Large Language Models modify code, inspect logs, revise plans, invoke tools, spend tokens, and decide whether objectives have been satisfied, affecting computational outcomes in ways that require accountable provenance. This paper introduces PROV-AGENT-LOOP, a W3C PROV-compatible model that represents objective-plan-act-evaluate cycles as queryable provenance. The model builds on PROV-AGENT, aligns with MAS, BDI, ReAct, Reflexion, and MAPE-K concepts, and captures loop structure, communication, evaluation, authorization, delegation, and invocation evidence. We implement PROV-AGENT-LOOP in Flowcept, integrating agentic, workflow, machine-learning, scheduler, and telemetry records in one Provenance Database. We validate the model on Frontier, where an AI agent adapts a Dask-based training campaign. The database answers eight cross-domain queries with 100% structural evidence completeness under OPL+DPL, more than double the workflow-only baseline, while Codex token usage and database footprint are reported as observed values rather than controlled overhead.
Back to AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC Archive Listing
Back to Full Workshop Archive Listing