How DeerFlow Revolutionizes Deep Research with a Multi‑Agent AI Architecture

Recently, ByteDance released DeerFlow, a new open‑source deep‑research project based on LangStack, on its official GitHub organization.

GitHub repository: https://github.com/bytedance/deer-flow

Official website: https://deerflow.tech/

Origin of the name Deer is an abbreviation of Deep Exploration and Efficient Research (D‑E‑E‑R). "Flow" reflects the LangGraph‑based workflow design.

DeerFlow offers features such as deep research, MCP integration, AI‑enhanced report editing, and podcast generation, demonstrated in a video walkthrough.

Demo Highlights

Video timestamps:

00:07 Add MCP

00:19 Start deep research

00:22 Human‑in‑the‑Loop

00:43 Generate rich report

00:53 AI‑enhanced report editing

01:08 Podcast generation

Multi‑Agent architecture demo: https://deerflow.tech/#multi-agent-architecture

Replay examples (e.g., Eiffel Tower vs tallest building): https://deerflow.tech/chat?replay=eiffel-tower-vs-tallest-building

Project Features

New Multi‑Agent Architecture

The custom Research Team mechanism supports multi‑turn dialogue, decision‑making, and task execution, reducing token consumption and API calls compared with the original LangChain Supervisor. A Re‑planning mechanism adds flexibility for dynamic task adaptation.

LangStack‑Based Open‑Source Framework

DeerFlow is built on LangChain ( https://github.com/langchain-ai/langchain ) and LangGraph ( https://github.com/langchain-ai/langgraph ), offering a clear code structure that lowers the learning curve for newcomers.

MCP Seamless Integration

Like Cursor and Claude Desktop, DeerFlow acts as an MCP host, allowing extensions of the Researcher Agent with private search, knowledge‑base access, and computer/phone/browser usage.

AI‑Generated Prompt (Meta Prompt)

All prompts are generated by OpenAI’s official Meta Prompt ( link ), ensuring high‑quality prompts and lowering prompt‑engineering barriers.

Human‑in‑the‑Loop

Users can modify and optimize AI‑generated plans or reports in real time via natural‑language instructions.

Podcast & PPT Generation

DeerFlow can create dual‑host podcasts using Volcano Engine’s speech technology ( link ) and generate PPTs from reports.

Core Tools

Leveraging LangChain’s extensive tool ecosystem, DeerFlow includes built‑in tools such as web search, code execution, and Jina Reader integration.

MCP Integration Example

async with MultiServerMCPClient({
    "math": {
        "command": "python",
        "args": ["/path/to/math_server.py"],
        "transport": "stdio",
    },
    "weather": {
        "url": "http://localhost:8000/sse",
        "transport": "sse",
    }
}) as client:
    all_tools = client.get_tools()
    ...

StateGraph and Handoffs

LangGraph’s StateGraph enables clear state‑machine design where each Agent is a node and edges define control flow. Handoffs allow one Agent to transfer control to another, implemented via Command objects.

def router(state) -> Command[Literal["researcher", "chat_generator"]]:
    goto = get_next_agent(...)
    return Command(goto=goto, update={"my_state_key": "my_state_value"})

DeerFlow’s architecture includes a Coordinator, Planner, Research Team (Researcher & Coder), Reporter, and optional Human Feedback loops, forming a robust pipeline from user query to final research report.

Core Tools Overview

DeerFlow bundles a variety of LangChain tools, illustrated in the diagram below.

Multi‑Agent Architecture Diagram

Podcast Generation Prompt Example

You are a professional podcast editor for a show called "Hello Deer." Transform raw content into a conversational podcast script suitable for two hosts to read aloud.

# Guidelines
- Tone: natural, conversational, with casual expressions.
- Hosts: one male, one female, alternating frequently.
- Length: ~10 minutes runtime.
- Output: JSON script without markdown.

DeerFlow invites contributions from developers, product managers, UX designers, and AI enthusiasts. Star the repository, submit commits, and share the project to help grow this open‑source AI legend.