About ZipperGen

Global protocols for inspectable agent systems.

ZipperGen is a Python DSL and runtime for structured multi-agent LLM coordination, based on the theory of Message Sequence Charts. You write one global protocol; ZipperGen projects it onto each agent and runs the result concurrently.

The Idea

Multi-agent frameworks often assemble workflows from callbacks, tool calls, and agent-local control flow. ZipperGen instead asks for a global protocol: who sends what, who runs which action, and who owns each decision.

That global protocol is then projected to one local program per lifeline. The result is executable, inspectable, and structurally deadlock-free for the coordination layer.

What You Get

  • A compact Python DSL for global coordination protocols
  • Automatic projection to per-agent local programs
  • Concurrent execution with FIFO message queues
  • Deadlock-freedom for the coordination layer, by construction
  • Human control points for approvals, reviews, corrections, and data entry
  • Causal runtime guards for stale-message scenarios
  • ZipperChat: a live message-sequence-chart viewer

Paper and Code

Research-backed, implementation-first.

The formal basis is described in Provable Coordination for LLM Agents via Message Sequence Charts. The repository contains the implementation, examples, tests, and links to the Lean formalization.

GitHub repository arXiv paper