Decoupled Intelligence: Multi-Agent LLMs for Traffic Simulation

Summary

This paper tackles a fundamental challenge in applying Large Language Models (LLMs) to complex simulation workflows: monolithic agent architectures struggle with the complexity of end-to-end tasks, leading to reasoning failures, parameter inconsistency, and poor state management. The authors propose a multi-agent collaborative framework that decouples the traffic simulation pipeline into five specialized roles—Planner, Builder, Demand, Runner, and Analyst—each responsible for a distinct phase of the simulation lifecycle.

The key innovation is a state-persistent Orchestrator built on the Model Context Protocol (MCP), which ensures seamless data handover and environmental consistency across distributed agent actions. This architecture enables a closed-loop refinement process where simulation outcomes are iteratively analyzed and optimized against user-defined Key Performance Indicators (KPIs). Experimental results from role ablation studies show that the multi-agent framework significantly outperforms single-agent baselines in both task success rates and parameter accuracy.

Key Contributions

• Multi-agent role decomposition: Decouples the traffic simulation pipeline into five specialized LLM agents (Planner, Builder, Demand, Runner, Analyst), each handling a distinct phase of the workflow
• MCP-based Orchestrator: Introduces a state-persistent coordination layer using the Model Context Protocol to maintain environmental consistency and enable seamless data handover between agents
• Closed-loop refinement: Implements iterative optimization where simulation outcomes are analyzed against user-defined KPIs and used to refine subsequent simulation runs
• Empirical validation: Demonstrates through role ablation studies that the multi-agent architecture achieves significantly higher task success rates and parameter accuracy compared to monolithic single-agent baselines
• Real-world applicability: Validates the system on real-world network extraction and traffic optimization tasks, showing the ability to bridge natural language intent with low-level simulation execution

Implications

For Researchers

This paper provides a concrete, validated architecture for multi-agent LLM systems in simulation domains. The role ablation study methodology offers a template for evaluating the contribution of each agent role to overall system performance. The MCP-based state management approach is a significant contribution to the challenge of maintaining context across distributed agent actions, which is a known bottleneck in multi-agent systems.

For Developers

The framework demonstrates how to implement MCP for state persistence in a multi-agent system, which is directly applicable to other complex workflows beyond traffic simulation. The specialized agent roles (Planner, Builder, etc.) provide a reusable pattern for decomposing any multi-step pipeline into discrete, manageable LLM tasks. The closed-loop refinement pattern is particularly valuable for any optimization or iterative improvement task.

For Users

For end users in urban planning and transportation analysis, this system lowers the barrier to entry for running complex traffic simulations. Users can express high-level goals in natural language, and the multi-agent system handles the technical details of scenario generation, execution, and optimization. This democratizes access to simulation tools that previously required significant technical expertise.

References

• https://arxiv.org/abs/2605.27685v1