Graph orchestration
Typed StateGraph with conditional edges, sub-graphs, recursion limits, and explicit routing. Agents you can read, log, and replay.
By 10xScale · v0.8 · MIT · Python 3.12+
Production AI agents in Python. Ship in minutes.
The open-source framework 10xScale built to power all their AI products — and open-sourced so every team could start from the same battle-tested foundation.
Typed graphs, durable threads, a REST and SSE server, and a typed TypeScript client. A modern alternative to LangGraph, CrewAI, and AutoGen — with the full production stack included.
bash
pip install 10xscale-agentflow 10xscale-agentflow-cli
agent.py
from agentflow.core.state import AgentState, Messagefrom agentflow.prebuilt.agent import ReactAgentdef get_weather(location: str,) -> str:return f"The weather in {location} is sunny."react_agent = ReactAgent(model="google/gemini-2.5-flash",system_prompt=[{"role": "system","content": "You are a helpful assistant. Use tools when they help answer the user.",}],tools=[get_weather],)app = react_agent.compile()result = app.invoke({"messages": [Message.text_message("Weather in Tokyo?")]},config={"thread_id": "demo", "recursion_limit": 10},)
Works with the stack you already use
OpenAI
Anthropic
Gemini
Vertex AI
Postgres
Redis
Qdrant
FastAPI
Next.js
Vercel
Docker
Python
TypeScript
Framework foundations
Everything a serious agent app needs.
One Python project. Typed graphs, durable threads, a production server, and a typed TypeScript client. Without the glue tax.
Production persistence
InMemoryCheckpointer for dev, PgCheckpointer (Postgres + Redis) for prod. Threads survive restarts and resume across replicas.
Backend to frontend
Built-in REST and SSE server, a typed TypeScript client, and a hosted playground. One project, full stack, no glue.
Built by 10xScale
The framework behind every 10xScale product.
10xScale is an AI-product company. AgentFlow is the runtime they built to power their products in production — not a demo, not a side project. When a feature was missing, they built it. When a pattern was painful, they fixed it.
The result is a framework shaped entirely by real production constraints: durable threads that survive restarts, typed graphs that stay readable at scale, and an API server that ships on day one rather than sprint ten. Then they open-sourced all of it so every team could start from the same foundation.
Visit 10xscale.ai →- Used in production at 10xScale across all products
- Typed StateGraph — explicit, readable, replayable
- Redis + Postgres dual-layer persistence
- Multi-model: OpenAI, Gemini, Anthropic, and custom
- REST + SSE server included — zero extra config
- Typed TypeScript client with React streaming hooks
- JWT auth, rate limiting, Snowflake IDs built in
Same agent, every framework
A familiar mental model, without the glue.
The same ReAct agent in five different frameworks. AgentFlow keeps the graph explicit and ships the production stack with it.
agent.py
from agentflow.core.state import AgentState, Messagefrom agentflow.prebuilt.agent import ReactAgentdef get_weather(location: str,tool_call_id: str | None = None,state: AgentState | None = None,) -> str:return f"The weather in {location} is sunny."react_agent = ReactAgent(model="google/gemini-2.5-flash",system_prompt=[{"role": "system", "content": "Helpful assistant."}],tools=[get_weather],trim_context=True,)app = react_agent.compile()
Documentation tracks
Pick the next useful page fast.
01
Start with a working Python agent
Install AgentFlow, run your first agent, and move from local code to a served app without guessing the next step.
02Learn the core concepts
Understand agents, tools, state, memory, streaming, and production runtime boundaries before you scale up.
03Use tutorials, how-to guides, and reference
Jump into example-driven tutorials, targeted implementation guides, and API reference for Python, REST, and TypeScript.
Connected stack
One project, four packages, zero glue.
The Python runtime, the API server, the TypeScript client, and the hosted playground are designed together. Same types, same threads, same primitives across the whole stack.
Beginner-friendly by design
A docs path that teaches the product.
Install the library, build one agent, add tools and memory, expose it through the API, connect a client, then deploy.
First agent
Create a working agent and understand the moving parts.
Tools and state
Give the agent capabilities; learn how state moves through a workflow.
Multi-agent flow
Compose agents into predictable handoffs and reusable workflows.
Production surface
Add persistence, APIs, streaming, clients, and deployment.
FAQ
Questions about AgentFlow.
Common questions about the framework, how it compares to alternatives, and how 10xScale uses it in production.
Frequently asked questions
What is AgentFlow?
AgentFlow (package: 10xscale-agentflow) is an open-source Python framework for building production-grade multi-agent AI systems. It provides a typed StateGraph for workflow orchestration, durable thread persistence via Redis and Postgres, a built-in REST and SSE API server, and a TypeScript client SDK. It is a modern, batteries-included alternative to LangGraph, CrewAI, AutoGen, and Google ADK.
Who built AgentFlow and why?
10xScale, an AI-product company, built AgentFlow to power their own production AI products. After finding that existing frameworks required too much custom plumbing to reach production, 10xScale built a batteries-included framework and open-sourced it under the MIT license. Every product at 10xScale runs on AgentFlow.
How is AgentFlow different from LangGraph?
AgentFlow and LangGraph share the graph-based mental model, but AgentFlow ships the full production stack: a built-in REST and SSE API server, a typed TypeScript client, a hosted playground, and dual-layer Redis + Postgres persistence. LangGraph requires separate LangServe or LangSmith integration and pulls in the LangChain dependency tree. AgentFlow has no LangChain dependency.
How is AgentFlow different from CrewAI or AutoGen?
CrewAI uses a role-based DSL that hides control flow; adding persistence and an API server requires significant extra glue. AutoGen is conversation-driven with LLM-powered selectors, making deterministic routing hard to reason about. AgentFlow keeps the graph explicit, typed, and readable, with production infrastructure included from day one.
Which AI models does AgentFlow support?
AgentFlow supports OpenAI (GPT-4o, o1, o3-mini), Google Gemini (Gemini 2.5 Flash, Gemini 2.0), Anthropic Claude (Claude 3.5, Claude 4), and any model exposed via a compatible API. You switch models by changing the model string — the graph and tools stay exactly the same.
Does AgentFlow support streaming?
Yes. AgentFlow ships a built-in SSE (Server-Sent Events) endpoint on the API server. The TypeScript client (@10xscale/agentflow-client) includes React hooks for streaming responses token-by-token or message-by-message. The hosted playground uses these hooks.
Can I use AgentFlow with TypeScript or a Next.js frontend?
Yes. The @10xscale/agentflow-client npm package is a fully typed TypeScript SDK that covers all API endpoints: invoke, stream, threads, memory, and file operations. It ships React hooks for streaming and is compatible with Next.js, Vite, Remix, and any React project.
Is AgentFlow production-ready?
AgentFlow is the runtime used in production by 10xScale for all their AI products. It ships with Postgres + Redis dual-layer persistence for durable threads, JWT authentication, rate limiting, Snowflake ID generation for distributed deployments, and Docker/Kubernetes build support via `agentflow build --docker-compose`.
How do I install AgentFlow?
Install the core library and CLI with: pip install 10xscale-agentflow 10xscale-agentflow-cli. Then scaffold a project with `agentflow init`, start a dev server with `agentflow api`, and open the playground with `agentflow play`. The TypeScript client is available via: npm install @10xscale/agentflow-client.
What is the license for AgentFlow?
AgentFlow is released under the MIT license. You can use it freely in commercial and open-source products without restriction. The full source code is available on GitHub at github.com/10xHub/Agentflow.
Ship a working agent in five minutes.
Install, build a graph, expose an API, connect a TypeScript client. Then keep going, without rewriting anything.