Hierarchical Server Architecture for Agentic Science
A new hierarchical, dynamic architecture and software system has been developed to discover computational resources across diverse cloud, edge, and HPC systems, specifically designed for agentic science workloads. The system enables concurrent, asynchronous negotiation, selection, and dispatch of requests using secretary agents. In testing, the agents probed 51 real and simulated providers across 7 categories, and performed 19,973 negotiation and 6,952 selection simulations, achieving 87.71% negotiation accuracy with selection costs comparable to traditional strategies. The architecture is designed for extensibility and supports multiple environments, addressing the need for automated resource discovery in scientific pipelines and workload managers. The paper is available on arXiv under the identifier 2608.05332.
Key facts
- The architecture is hierarchical and dynamic.
- It enables concurrent, asynchronous negotiation, selection, and dispatch of requests.
- Secretary agents are used for resource discovery.
- 51 real and simulated providers across 7 categories were probed.
- 19,973 negotiation and 6,952 selection simulations were performed.
- Negotiation accuracy was 87.71%.
- Selection costs are comparable to traditional strategies.
- The system is designed for extensibility.
- The paper is on arXiv (2608.05332).
Entities
Institutions
- arXiv