Sochat's AI dispatch agent hits 97.9% success on 432 runs

Dispatch remains a bottleneck in multi-cluster scientific computing: translating a user request into a workload manager's job spec is error-prone, especially across heterogeneous clusters. In 'Descriptive Dispatch of Computational Work', Vanessa Sochat and Daniel Milroy from Sandia National Laboratories propose using AI/ML agents to handle this task autonomously. The agent receives plain-language work instructions, prepares the job specification, and submits it to the scheduler, effectively bridging the gap between human intent and cluster mechanics.

Across 432 test runs spanning five feature dimensions and four prompt styles, the agent achieved a 97.9% success rate, demonstrating robust reliability. But the most striking result comes from a full orchestration test: when 220 jobs were submitted across multiple clusters, adding descriptive metadata to the dispatch process raised successful execution from 48% to 87%. That metadata also eliminated architecture mismatches (e.g., requesting the wrong node type) and improved performance for 5 of 10 measurable applications by up to 3.3x. The authors' breakdown suggests that simply telling the agent what the job needs—rather than letting it guess—reduces costly failures and accelerates overall throughput. For teams wrestling with heterogeneous HPC resources, this points to a future where agents and descriptive specs become standard practice in scientific workflow orchestration.