Provable Coordination for LLM Agents via Message Sequence Charts

arXiv cs.AI · 6 hours ago

Researchers developed a domain-specific language based on message sequence charts for specifying coordination between multiple LLM agents, with a syntax-directed projection that generates deadlock-free local programs from global specifications. The framework, called ZipperGen, separates message-passing structure from unpredictable LLM and tool calls, establishing coordination properties independently of LLM nondeterminism. This enables multi-agent LLM systems to guarantee coordination correctness even when individual agent outputs remain unpredictable.