Agentic AI for High Performance Computing and Large Scale Science: A Narrative Review of Autonomous Orchestration, Coordination Challenges, and Evaluation Gaps

Abstract

Agentic artificial intelligence, systems built on large language models (LLMs) that autonomously plan, invoke tools, and act in multi-step loops, is increasingly being proposed as a control layer for high-performance computing (HPC) and large-scale scientific workflows. Agents are now used to request compute allocations, launch simulation ensembles, steer experiment campaigns, move petabyte-scale datasets, and coordinate hybrid AI-simulation pipelines across the edge-cloud-HPC continuum. This narrative review synthesizes a fast-growing but fragmented body of work spanning distributed and parallel computing (DC), multi-agent systems (MA), and computational science, to answer three questions: (1) what architectural patterns are being used to connect agentic reasoning to HPC and scientific execution substrates; (2) what parallel-computing-style coordination and scheduling problems arise when agents act concurrently at scale; and (3) how rigorously are these systems currently evaluated. We propose a taxonomy of agent roles (resource-request, workflow-steering, data-movement, and co-scheduling agents), review representative systems and case studies in a comparative table, and synthesize cross-cutting open problems in hallucination propagation, scheduling visibility, energy accounting, and reproducibility. We argue that the field's central weakness is not architectural but methodological: evaluation of agentic HPC systems remains almost entirely non-standardized, with few shared benchmarks, workloads, or cost metrics across studies. We close with a research agenda intended to guide the next generation of parallel-computing contributions to agentic AI for science, and we explicitly scope this as a narrative rather than a PRISMA-style systematic review.