Introduction
"Moving from 'chatting' to 'executing', agents need a full 'harness' to provide both constraint and empowerment."
This is the 68th article in the "One Open Source Project a Day" series. Today, we explore DeerFlow.
If you've followed OpenAI's Deep Research or similar advanced search tools, you'll be impressed by ByteDance's open-source DeerFlow. Initially gaining fame as a deep research tool, it has evolved into a fully functional Super Agent Harness in its 2.0 version. It doesn't just search for information; it writes code in isolated sandboxes, runs experiments, generates reports, and even collaborates with teams via Slack or Feishu/Lark.
What You Will Learn
- How DeerFlow evolved from deep search to a universal Super Agent foundation.
- Core architecture: Multi-agent orchestration and state management based on LangGraph.
- Security: Safely executing AI-generated code in Docker/K8s sandboxes.
- How the Long-term Memory system makes agents smarter over time.
- Deploying dedicated researcher agents within enterprise platforms (Lark, Slack).
Prerequisites
- Basic understanding of Large Language Models (LLMs).
- Familiarity with the Python development environment.
- Preliminary knowledge of Agents and RAG (Retrieval-Augmented Generation).
Project Background
Overview
DeerFlow (Deep Exploration and Efficient Research Flow) is a deep research and task execution framework developed and open-sourced by ByteDance's Volcengine/BytePlus teams. It is designed to solve the stability issues agents face during long-running (minutes to hours) tasks with complex logic. Through modular Skill definitions, strict sandbox isolation, and persistent memory, DeerFlow enables AI to truly take on roles like "Researcher" or "Assistant."
Author/Team Introduction
- Team: ByteDance (Volcengine / BytePlus).
- Evolution: Version 2.0 was released in February 2026, completely refactored as an agentic foundation.
- Accolades: Has topped GitHub Trending multiple times, sparking widespread developer discussion.
Project Data
- ⭐ GitHub Stars: 59.4k+
- 🍴 Forks: 7.5k+
- 📄 License: MIT License
- 🌐 GitHub: https://github.com/bytedance/deer-flow
Key Features
Core Value
DeerFlow's core value lies in providing a "batteries-included" agent runtime. It provides LLMs with file systems, execution environments (sandboxes), long-term memory, and multi-channel communication, allowing them to handle complex tasks requiring deep thinking and iterative experimentation.
Use Cases
-
Deep Industry Research
- Automatically search, filter, and synthesize global web information to generate professional, multi-thousand-word research reports.
-
Automated Content Creation
- Generate professional slide decks (PPTs), technical documentation, or static websites based on research results.
-
Data Pipelines & Code Execution
- AI writes and runs Python scripts in sandboxes for data processing, keeping the host system secure.
-
Enterprise Assistant Agents
- Integrate into Lark or Slack to provide automated weekly report summaries, competitive monitoring, and more.
Quick Start
You can quickly invoke it via the Embedded Client in a Python script:
from deerflow.client import DeerFlowClient
# Initialize the client
client = DeerFlowClient()
# 1. Start a simple research task
response = client.chat("Research the latest commercialization progress of solid-state batteries in 2026", thread_id="research-01")
print(response["content"])
# 2. Stream task progress for long-running tasks
for event in client.stream("Analyze the current competitive landscape of Generative AI Video"):
if event.type == "messages-tuple" and event.data.get("type") == "ai":
print(event.data["content"], end="", flush=True)Core Characteristics
- Complex Orchestration via LangGraph: Supports task decomposition, parallel execution, reflection, and recovery in cyclic workflows.
- Markdown-Defined Skills: Developers can define complex agent behaviors by writing Markdown, lowering the entry barrier.
- Isolated Sandbox (AioSandbox): Supports Docker or K8s containers to safely execute AI-generated code.
- Long-Term Memory System: Persistent storage of user preferences, style, and facts for personalized intelligence.
- Multi-Channel Integration: Built-in integration for Telegram, Slack, Feishu/Lark, and WeChat.
Project Advantages
| Dimension | DeerFlow | Ordinary Chatbot / RAG |
|---|---|---|
| Task Duration | Supports long-horizon (mins/hours) tasks | Mainly supports instant responses (secs) |
| Capabilities | File Read/Write, Code Run, Message Send | Primarily text generation |
| Security | Enforced sandbox isolation | Usually no isolation or local only |
| Memory Depth | Persistent cross-session refinement | Limited to current conversation context |
Why choose this project?
- ByteDance Backing: Validated at scale for engineering reliability and performance.
- Complete Ecosystem: Provides everything from frontend UI to backend sandboxes and IM integrations.
Detailed Analysis
Architecture: The Agentic "Control Tower"
DeerFlow's architecture embodies the essence of Harness Engineering:
1. Workflow Orchestration (LangGraph)
The heart of DeerFlow. Using Directed Acyclic Graphs (DAG) or graphs with cycles, it decomposes complex tasks into nodes like Planning -> Search -> Execution -> Review -> Reporting.
2. Skill & Tool Loading Mechanism
Business logic is encapsulated into Skills. The AI only loads relevant Skill contexts when needed, significantly saving Tokens and improving instruction-following accuracy.
3. Sandbox Isolation (AioSandbox Provider)
When an agent needs to calculate or process files, DeerFlow dynamically creates an isolated container. Even if AI-generated code contains malicious commands like rm -rf /, it cannot harm the host due to sandbox restrictions—critical for enterprise deployments.
4. Memory Closure (Memory Loop)
Beyond storing chat logs, DeerFlow automatically extracts "knowledge" and "habits" from conversations, saving them into a long-term memory store for automatic context completion in future tasks.
Project Resources
Official Resources
- 🌟 GitHub: https://github.com/bytedance/deer-flow
- 📚 Docs: Official Docs
- 💬 Community: GitHub Issues / ByteDance internal discussion groups.
Target Audience
- AI Developers: Building Super Agents with complex logic and productivity tools.
- Industry Analysts: Needing automated, deep-dive research tools.
- Enterprise Digital Teams: Seeking secure, controllable AI assistant deployment.
Welcome to visit my Homepage to find more useful knowledge and interesting products.