New DSL for LLM Agent Coordination Using Message Sequence Charts

A new domain-specific language (DSL) has been developed by researchers to define coordination among agents based on large language models (LLMs) through message sequence charts (MSCs). This language distinguishes the structure of message passing from the unpredictable actions of LLMs, facilitating deadlock-free local agent programs through syntax-directed projection. The introduction of a diagnosis consensus protocol demonstrates the establishment of coordination properties independent of LLM nondeterminism. Additionally, a runtime planning extension enables an LLM to create coordination workflows dynamically while maintaining the same structural guarantees. An open-source implementation in Python is available.