Towards Multi-Agent Autonomous Reasoning in Hydrodynamics

A new preprint on arXiv introduces a multi-agent autonomous reasoning system tailored for hydrodynamics, tackling a key limitation of single-agent LLM workflows: context saturation. As tool specifications and traces accumulate, a single agent's effective reasoning context shrinks, degrading reliability. The authors — Jinpai Zhao, Albert Cerrone, Joannes Westerink, and Clint Dawson — present a Multi-Agent System (MAS) prototype built around a Layer Execution Graph (LEG) that coordinates specialized agents via natural-language routing heuristics.

The architecture consists of a planner agent that constructs query-specific execution topologies from domain heuristics, specialist agents that operate under strict tool allowlists in separate data roles, consolidator agents that fuse parallel outputs into concise briefs, and a reporter agent that synthesizes the final response. All agents use Claude Sonnet 4.6 as the backbone. The runtime logs provenance for every tool invocation to support auditability — critical for scientific applications.

Evaluated on 37 queries spanning six complexity categories, the system achieves 93.6% factual precision with a 100% pass rate. Accuracy remains above 90% across runs from single-threaded to five independent parallel tracks. Under simulated loss of individual data sources, the system degrades gracefully, still returning substantive partial answers. These results suggest that planner-guided, graph-structured multi-agent orchestration can significantly alleviate context-saturation bottlenecks that constrain monolithic single-agent architectures, making this approach promising for other scientific domains as well.