Orchestrated multi agents sustain accuracy under clinical-scale workloads compared to a single agent
Abstract We tested state-of-the-art LLMs under clinical-scale workloads using two designs: a single agent handling all tasks and a multi-agent orchestrator assigning each task to a dedicated worker. Across retrieval, extraction, and dosing tasks, batch sizes ranged from 5–80. Multi-agent accuracy remained high (90.6% at 5 tasks; 65.3% at 80), while single-agent accuracy collapsed (73.1% to 16.6%; p < 0.01). Multi-agent runs used up to 65-fold fewer tokens and limited latency growth. These findings show that lightweight orchestration preserves accuracy and efficiency under mixed-task clinical loads.
Authors
- Mahmud Omar
- Prem Timsina (ORCID: https://orcid.org/0000-0002-6047-887X)
- Nicholas Gavin (ORCID: https://orcid.org/0000-0001-6900-3953)
- Alexander W. Charney (ORCID: https://orcid.org/0000-0001-8135-6858)
- Lisa Stump (ORCID: https://orcid.org/0000-0002-1327-4110)
- Eyal Klang
- Girish N. Nadkarni
- Benjamin S. Glicksberg
- Ganesh Raut
- Robert Freeman
- Reem Agbareia
Institutions
- Hadassah Medical Center (IL)
- Mount Sinai Medical Center (US)
- Child Health and Development Institute (US)
- Pediatrics and Genetics (US)
- Icahn School of Medicine at Mount Sinai (US)
Publication Details
- Journal
- npj Health Systems
- Published
- 2026-03-09
- DOI
- https://doi.org/10.1038/s44401-026-00077-0
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Georgia Clinical and Translational Science Alliance
- National Institutes of Health