About ZipperGen
ZipperGen is a Python DSL and runtime for structured multi-agent LLM coordination, grounded in the theory of Message Sequence Charts (MSCs). You write a single global protocol; ZipperGen projects it onto each agent and runs them concurrently, with deadlock-freedom guaranteed by construction.
The idea
Multi-agent frameworks typically assemble pipelines from callbacks and tool calls, with no global view of coordination and no formal semantics. ZipperGen takes a different approach: you describe the entire choreography once, as a global workflow, and the framework automatically derives the local program for each agent. The theoretical foundation comes from the study of Message Sequence Charts and communicating automata — a well-developed area of formal methods.
What you get
- A clean Python DSL for writing global coordination protocols
- Automatic projection onto per-agent local programs
- Concurrent execution with FIFO message queues
- Deadlock-freedom for the coordination layer, by construction
- Built-in backends for OpenAI, Mistral, and Claude
- ZipperChat: a live message-sequence-chart viewer in your browser
The team
ZipperGen is developed by Benedikt Bollig, Matthias Függer, and Thomas Nowak at Université Paris-Saclay, CNRS, ENS Paris-Saclay, and LMF. It accompanies the paper Provable Coordination for LLM Agents via Message Sequence Charts.