Multi-Agent System for Hydrodynamic Simulations Using LLMs
A prototype of a multi-agent system (MAS) for hydrodynamics has been developed, replacing traditional single-agent LLM processes with specialized agents managed via a Layer Execution Graph (LEG). A planner agent creates execution topologies tailored to specific queries using natural-language routing heuristics, thus incorporating domain expertise without inflexible control mechanisms. Specialist agents function under defined tool allowlists and fulfill distinct data-class roles. Meanwhile, consolidator agents merge outputs from various layers, and a reporter agent compiles the final response. Additionally, runtime logs maintain provenance. This innovative method tackles problems related to context window reduction and the reliability of single-agent scientific workflows.
Key facts
- Single-agent systems (SAS) are the default for LLM-driven scientific workflows but suffer from context shrinkage and reliability issues.
- The multi-agent system (MAS) prototype is designed for hydrodynamics.
- Specialized agents are coordinated through a Layer Execution Graph (LEG).
- A planner agent constructs query-specific execution topologies from natural-language routing heuristics.
- Specialist agents operate under strict tool allowlists and occupy complementary data-class roles.
- Consolidator agents fuse parallel outputs into concise briefs between layers.
- A reporter agent synthesizes the final response.
- The runtime logs provenance.
Entities
—