duragraph 0.3.3

Python SDK for DuraGraph — durable, replayable agent workflows

DuraGraph Python SDK

Python SDK for DuraGraph — durable, replayable agent workflows.

Define a graph with decorators, register it with the control plane, and get crash-safe execution, event-sourced replay, and live observability for free. Like Temporal for AI agents — in Python.

Installation

The package is published as duragraph on PyPI (renamed from duragraph-python in v0.3.0; the old name is frozen at 0.2.1 and will not receive further releases). We recommend uv for dependency management.

# With uv (recommended)
uv add duragraph
uv add 'duragraph[openai]'
uv add 'duragraph[anthropic]'
uv add 'duragraph[all]'

# With pip
pip install duragraph

Quick Start

from duragraph import Graph, llm_node, entrypoint

@Graph(id="customer_support")
class CustomerSupportAgent:
    """A customer support agent that classifies and responds to queries."""

    @entrypoint
    @llm_node(model="gpt-4o-mini")
    def classify(self, state):
        """Classify the customer intent."""
        return {"intent": "billing"}

    @llm_node(model="gpt-4o-mini")
    def respond(self, state):
        """Generate a response based on intent."""
        return {"response": f"I'll help you with {state['intent']}."}

    # Define flow
    classify >> respond


# Run locally
agent = CustomerSupportAgent()
result = agent.run({"message": "I have a billing question"})
print(result)

# Or deploy to control plane
agent.serve("http://localhost:8081")

Features

Decorator-Based Graph Definition

from duragraph import Graph, llm_node, tool_node, router_node, human_node

@Graph(id="my_agent")
class MyAgent:
    @llm_node(model="gpt-4o-mini", temperature=0.7)
    def process(self, state):
        return state

    @tool_node
    def search(self, state):
        results = my_search_function(state["query"])
        return {"results": results}

    @router_node
    def route(self, state):
        return "path_a" if state["condition"] else "path_b"

    @human_node(prompt="Please review")
    def review(self, state):
        return state

Streaming

async for event in agent.stream({"message": "Hello"}):
    if event.type == "token":
        print(event.data, end="")
    elif event.type == "node_completed":
        print(f"\nNode {event.node_id} completed")

Subgraphs

@Graph(id="research")
class ResearchAgent:
    @llm_node
    def research(self, state):
        return {"findings": "..."}

@Graph(id="main")
class MainAgent:
    research = ResearchAgent.as_subgraph()

    @entrypoint
    def plan(self, state):
        return state

    plan >> research

Requirements

• Python 3.10+
• DuraGraph Control Plane (for deployment)

Documentation

• Full Documentation
• API Reference
• Examples (examples/)

Related modules

The Python SDK lives in the duragraph monorepo alongside its siblings:

Path	Description
/	Core engine (Go, embeds the React dashboard incl. visual workflow editor)
go-sdk/	Go SDK
examples/	Example projects (multi-language)
docs/	Documentation site source

Contributing

See CONTRIBUTING.md for guidelines.

License

Apache 2.0 - See LICENSE for details.