Deadlock-free LLM agent coordination

ZipperGen is a Python DSL and runtime for structured multi-agent LLM coordination. You write a single global protocol describing what messages flow between which agents and who owns each decision. ZipperGen projects it onto each agent and runs them concurrently, with deadlock-freedom guaranteed by construction. The guarantees are formally established in our paper.

Quick start

git clone https://github.com/zippergen-io/zippergen.git
cd zippergen
pip install -e .

Python 3.11 or later required.

Hello, World!

Three agents collaborate: Writer drafts a tweet, Editor decides whether it's good enough, and ZipperGen handles the coordination. No API key needed; the built-in mock backend runs instantly. 

from zippergen.syntax import Lifeline, Var
from zippergen.actions import llm
from zippergen.builder import workflow

User   = Lifeline("User")
Writer = Lifeline("Writer")
Editor = Lifeline("Editor")

topic    = Var("topic",    str)
tweet    = Var("tweet",    str)
approved = Var("approved", bool)

@llm(system="Write a one-sentence tweet about the topic.",
     user="{topic}", parse="text", outputs=(("tweet", str),))
def draft(topic: str) -> None: ...

@llm(system="Is this tweet engaging, original, and under 180 chars? Reply true or false.",
     user="{tweet}", parse="bool", outputs=(("approved", bool),))
def approve(tweet: str) -> None: ...

@llm(system="Improve this tweet: shorter and punchier.",
     user="{tweet}", parse="text", outputs=(("tweet", str),))
def revise(tweet: str) -> None: ...

@workflow
def write_tweet(topic: str @ User) -> str:
    User(topic) >> Writer(topic)
    Writer: tweet = draft(topic)
    Writer(tweet) >> Editor(tweet)
    Editor: approved = approve(tweet)
    if approved @ Editor:
        Editor(tweet) >> User(tweet)
    else:
        Editor(tweet) >> Writer(tweet)
        Writer: tweet = revise(tweet)
        Writer(tweet) >> User(tweet)
    return tweet @ User

# No API key needed; runs with the built-in mock backend.
# Switch to a real LLM: write_tweet.configure(llms="openai")
result = write_tweet(topic="a git commit message that tells the truth")
print(result)

Using real LLMs

Set your API key, pass a provider name, and run: 

export OPENAI_API_KEY=...
my_workflow.configure(llms="openai", ui=True)
result = my_workflow(notes="...", diagnosis="...")

Built-in providers: openai, mistral, claude. Different agents can use different providers: 

my_workflow.configure(llms={"LLM1": "mistral", "LLM2": "openai"})

ZipperChat

Pass ui=True to open a live message-sequence-chart in your browser as the agents run. Watch who sends what, who is thinking, and where the workflow currently is.

ZipperChat showing a write_tweet workflow execution as a live message sequence chart

Formal foundation

ZipperGen is grounded in the theory of Message Sequence Charts. The projection from a global protocol to per-agent local programs is syntax-directed, and deadlock-freedom follows by structural induction; no runtime checking required. The formal proofs are in our paper:

Bollig, Függer, Nowak. Provable Coordination for LLM Agents via Message Sequence Charts. arXiv:2604.17612 [cs.PL]

Learn more