AgentSUMO: An Agentic Framework for Interactive Simulation Scenario Generation in SUMO via Large Language Models¶

🕒 Published (v1): 2025-11-10 07:46 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

AgentSUMO is an LLM-driven agentic framework that converts abstract, high-level traffic policy goals (e.g., "reduce congestion near Gangnam Station") into fully executable SUMO simulation workflows without requiring expert knowledge from the user. It pairs an Interactive Planning Protocol (IPP) for adaptive multi-step reasoning with the Model Context Protocol (MCP) for standardized tool orchestration. Experiments on Seoul and Manhattan urban networks demonstrate substantial improvements in traffic flow metrics over imperative baselines.

Problem¶

Existing LLM-integrated SUMO systems (ChatSUMO, SUMO-MCP) remain locked in imperative workflows: users must still decompose their goals into explicit procedural commands ("generate network," "configure signals," "run simulation"), presupposing domain expertise that non-expert policymakers and urban planners do not possess. Abstract, incomplete user intent—the norm for non-experts—is structurally incompatible with these command-driven pipelines. ChatSUMO provides conversational accessibility but rigid step sequencing; SUMO-MCP provides protocol-driven flexibility but lacks an interactive dialogue loop. Neither supports goal-oriented, end-to-end autonomous planning.

Method¶

AgentSUMO is organized around two orthogonal layers:

Interactive Planning Protocol (IPP) — embedded in the Planner Agent's system prompt via prompt engineering: 1. Task Complexity Assessment: classifies each query as Simple (direct execution), Complex (Chain-of-Thought within <thinking> tags), or Agentic (multi-phase extended CoT covering solution-space exploration, trade-off analysis, and synthesis). 2. Parameter Sufficiency Validation: maps each request to a scenario schema (spatial scope, OD data, policy interventions, evaluation metrics); pauses and elicits structured clarification if any required field is missing. 3. Clarify-before-Execute: the agent summarizes the full intended plan and waits for explicit user confirmation before invoking any tools; simple, fully-specified tasks proceed directly.

Model Context Protocol (MCP) — three-tier host/client/server architecture (following Anthropic's MCP spec, inspired by LSP) that decouples tool implementation from usage via dynamic schema discovery and bi-directional communication. Two MCP servers are deployed: the AgentSUMO MCP Server (scenario generation, routing, policy experimentation, visualization) and an SQLite MCP Server (structured result storage enabling cross-scenario queries).

Supporting infrastructure: multi-turn conversation history embedding tool results in the message sequence; a simulation state dictionary (XML-injected into each system prompt) tracking generated network/route/output artifacts to prevent redundant re-computation; and prompt caching on static system-prompt sections to reduce per-turn token overhead.

Key Contributions¶

End-to-end agentic SUMO orchestration framework (AgentSUMO) enabling goal-oriented natural-language-to-simulation translation for non-expert users.
Interactive Planning Protocol: adaptive reasoning depth scaled to task complexity (Simple/Complex/Agentic) via structured prompt logic—no fine-tuning required.
MCP-based tool layer providing standardized, dynamically discoverable simulation tool integration with auditable access control.
State-aware multi-turn session management that supports progressive, cumulative scenario refinement without manual file handling.
Prompt caching optimization for production-scale multi-turn deployment efficiency.
Empirical evaluation on real urban networks (Seoul, Manhattan) across policy scenarios.

Results¶

AgentSUMO achieves "substantial improvements in traffic flow metrics" on Seoul and Manhattan urban networks compared to imperative baselines (specific metric values not provided in the supplied text excerpt).
The agentic workflow successfully handles multi-level policy experimentation (lane adjustments, signal timing, OD changes) while remaining accessible to non-expert users without predefined scripting.
The framework is validated against ChatSUMO (pipeline-based) and SUMO-MCP (protocol-only) as the primary baselines.

(Note: Quantitative tables appear beyond the supplied text; only qualitative outcome claims are available above.)

Limitations¶

No quantitative ablation of IPP components (e.g., cost/accuracy tradeoff of Agentic vs. Complex CoT paths) is visible in the provided text.
Prompt caching and IPP effectiveness depend on static system-prompt structure; highly dynamic tool schemas would invalidate cache segments.
Evaluation is limited to two urban networks (Seoul, Manhattan); generalizability to other simulator topologies or edge-case policy configurations is unassessed.
The Clarify-before-Execute loop introduces latency per session; for real-time decision support the turn-round-trip cost is not analyzed.
Reliance on Anthropic's Claude as the backbone LLM (MCP is Anthropic's spec); portability to open-weight models is not evaluated.

Relevance to Harnesses / Meta-Harnesses¶

AgentSUMO is a textbook domain-specific meta-harness: it does not perform simulation itself but orchestrates a heterogeneous tool fleet (network generator, OD configurator, signal optimizer, visualizer, SQLite result store) through a protocol-standardized interface (MCP), while a planning agent decides which tools to invoke, in what order, and with what parameters—all derived from a single abstract user intent. The Interactive Planning Protocol is architecturally equivalent to a meta-harness routing layer that classifies incoming tasks and dispatches them to different reasoning sub-harnesses (direct, CoT, extended CoT). The pattern of injecting state context into each system prompt as XML and caching static prompt sections is directly transferable to any harness managing stateful multi-step pipelines over external tool APIs. For researchers tracking harness design, AgentSUMO demonstrates how MCP can serve as the standardized inter-component bus in a harness, and how prompt-engineering alone (without fine-tuning) can implement adaptive workflow control.