Provable Coordination for LLM Agents via Message Sequence Charts

2026-04-19

AI Summary

This work addresses the challenge of coordination in large language model (LLM)-based multi-agent systems, where nondeterministic LLM behavior can lead to subtle, hard-to-detect errors such as deadlocks or message mismatches. The paper introduces ZipperGen, a novel framework that formally incorporates Message Sequence Charts (MSCs) into LLM-driven multi-agent coordination for the first time. It employs a domain-specific language to decouple communication structure from LLM behavior and uses syntax-guided projection to derive local agent programs from a global specification, guaranteeing deadlock freedom by construction. This approach enables a verifiable coordination mechanism that is disentangled from LLM nondeterminism and supports runtime generation of structurally sound workflows. The framework’s ability to independently verify coordination properties is demonstrated through its application to consensus protocol diagnostics.

Technology Category

Multiagent Systems: Coordination and CollaborationMachine Learning: Large Multimodal Models (LMMs)Natural Language Processing: (Large) Language Models

Application Category

Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactionsEconomics, Online Markets and Human Computation: Cost models of using LLMs in production systemsUser Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Multi-agent systems built on large language models (LLMs) are difficult to reason about. Coordination errors such as deadlocks or type-mismatched messages are often hard to detect through testing. We introduce a domain-specific language for specifying agent coordination based on message sequence charts (MSCs). The language separates message-passing structure from LLM actions, whose outputs remain unpredictable. We define the syntax and semantics of the language and present a syntax-directed projection that generates deadlock-free local agent programs from global coordination specifications. We illustrate the approach with a diagnosis consensus protocol and show how coordination properties can be established independently of LLM nondeterminism. We also describe a runtime planning extension in which an LLM dynamically generates a coordination workflow for which the same structural guarantees apply. An open-source Python implementation of our framework is available as ZipperGen.

Problem

• multi-agent systems
• large language models
• coordination errors
• deadlocks
• message sequence charts

Innovation

• message sequence charts
• multi-agent coordination
• domain-specific language
• deadlock-free synthesis
• LLM agent systems