ECAS: An Edge-Controlled Agentic System for Validation-Gated Scientific Application Execution

Details

LLM agents can automate scientific computing workflows, but two obstacles block deployment: giving cloud-hosted models direct HPC access exposes credentials and execution authority; and one-shot code generation fails when artifacts encounter site-specific environment issues at scale.\n\nECAS (Edge-Controlled Agentic System) separates three concerns: a cloud LLM proposes plans and repairs; a user-controlled edge agent holds credentials, enforces policy, and gates execution; the HPC cluster computes. Generated artifacts pass static checks and small-scale validation before target-scale runs are permitted — failures trigger repair loops from sanitized feedback.\n\nExperiments on three scientific applications across two production ALCF systems (Argonne Leadership Computing Facility) with six injected fault types show: closed-loop repair improves success from 0/6 to 6/6 versus one-shot generation; validation gating prevents all three observed large-scale failures; edge-resident skill conditioning improves success from 4/6 to 6/6.\n\nThe architecture is directly applicable to enterprise HPC environments where security requirements prevent cloud model access to production systems.