Birds of a Feather Archive

Agentic Computing and Services for the Computing Continuum


Authors: Vipin Chaudhary (Case Western Reserve University), Dhabaleswar Panda (The Ohio State University)

Abstract: AI applications increasingly span a computing continuum from edge sensors through cloud to HPC centers, yet most are still assembled and operated manually. This BoF convenes the SC community around agentic, service-oriented cyberinfrastructure: software agents, exposed as composable services, that sense continuum state, decide where to train and deploy models, act through a shared control plane, and learn from outcomes. Drawing on experience from the NSF ICICLE AI Institute and invited external perspectives, participants will debate autonomy levels, trust, provenance, evaluation, and emerging protocols such as MCP, and identify steps toward community benchmarks and interoperable agentic cyberinfrastructure.

Long Description: Topic: Modern AI workloads are distributed across a computing continuum: data are born at the edge (sensors, drones, cameras), models are trained on cloud or HPC resources, and inference must run wherever latency, privacy, and cost dictate. Today these pipelines are wired together manually, binding models, data, and placement to one environment and one team of experts. A growing body of work argues that the next phase of cyberinfrastructure (CI) should be service-oriented (models, data, training, inference, and placement as uniform, composable services), agentic (software agents closing a sense-decide-act-learn loop over those services, with humans in or on the loop), and continuum-aware (decisions that account for the full edge-cloud-HPC span). The NSF ICICLE AI Institute provides an evidence base: a Tapis control plane, a Model and Data Commons, Patra Model Cards for provenance, CKN edge telemetry, and an early agentic MLOps demonstration. Open questions remain around decision verification, evaluation benchmarks, cost- and energy-aware autonomy, federated cross-institution operation, and human oversight.

Goals: The BoF has three goals: (1) surface community requirements and concerns for agentic operation of shared HPC, cloud, and edge resources, including autonomy levels appropriate for production facilities; (2) compare approaches to trust, provenance, policy-gated action, and verification of agent decisions; and (3) seed a community effort on evaluation, benchmarks, and interoperability contracts (including protocols such as MCP) for agentic CI.

Relevance to the SC audience: This topic sits at the intersection of AI/ML, system software, clouds and distributed computing, and state of the practice, all core SC concerns. HPC center operators face immediate questions about whether and how agents should schedule, retrain, and redeploy on their systems; tool and middleware developers need shared contracts; domain scientists want AI capabilities without operating pipelines by hand; and facility policy staff must weigh autonomy against accountability. Recent results in learned scheduling, LLM-assisted compilation, performance prediction, and log anomaly detection make the discussion timely rather than speculative.

Structure and speakers: To ensure the discussion is not vendor-specific or single-project-centric, short position statements will come from the session leaders and from three confirmed external technical speakers spanning the BoF's three themes: Ian Foster (Argonne National Laboratory and University of Chicago) on agentic AI for science and autonomous research workflows; Joe Stubbs (TACC) on service computing and AI across the edge-cloud continuum; and Varun Chandola (University at Buffalo, SUNY) on the computing continuum and AI cyberinfrastructure. Most of the session is reserved for moderated discussion.

Expected outcome: Concrete outputs include: a summarized list of community requirements, risks, and adoption barriers; an inventory of existing building blocks and gaps for agentic CI; candidate dimensions for a community evaluation effort (task success, resource use, recovery from failure, quality of human oversight); and a contact list of volunteers for a follow-on working group and a report to be shared openly with attendees and the broader community.


Back to Birds of a Feather Archive Listing