DeerFlow: ByteDance’s Open‑Source Super‑Agent That Executes Whole Projects End‑to‑End

Project Overview

DeerFlow is an open‑source super‑agent framework released by ByteDance in early 2026. Its core idea is a single instruction that yields end‑to‑end delivery. Unlike typical tool‑calling agents such as Hermes or CrewAI, DeerFlow builds a multi‑agent collaboration system that automatically decomposes tasks, launches sub‑agents, writes code in a sandbox, iteratively tests and self‑corrects, and finally produces ready‑to‑use artifacts.

Version 2.0, released March 2026, is a complete rewrite under the MIT license and does not inherit code from the 1.x line.

Core Architecture

Sub‑Agent Orchestration

DeerFlow ships with a hierarchy of role‑specific sub‑agents (planning, coding, testing, searching, etc.). A Message Gateway coordinates their communication, emulating a small development team.

Sandbox and File System

All code generation, execution and testing occur inside an isolated sandbox that can be deployed with Docker, keeping the host environment safe. After a task finishes, the sandbox retains the full work context.

Long‑Term Memory

The built‑in Memory module persists context across tasks, enabling reuse of previous results and reducing duplicated effort.

Skill Extensions

DeerFlow’s extensible Skill mechanism lets it connect to various toolchains, MCP servers, IM channels and other services, adapting to different workflows.

Main Feature Highlights

End‑to‑End Task Execution : Directly delivers usable outputs such as research reports, full websites, dashboards or slide decks without intermediate drafts.

Multi‑Model Support : Recommended models include Doubao Seed‑2.0‑Code, DeepSeek V3.2 and Kimi 2.5; other providers like OpenRouter, Groq or NVIDIA NIM can also be used.

Claude Code / Cursor Integration : Seamlessly invokes popular coding agents for users with existing local development environments.

Instant Search & Crawling : Bundles BytePlus’s InfoQuest for free online intelligent search and web‑scraping.

Flexible Deployment : Offers one‑click Docker deployment and a local development mode; Docker is the recommended default and can be set up in about five minutes.

Quick Start

Option 1 – Claude Code / Cursor (simplest)

Help me clone DeerFlow if needed, then bootstrap it for local development by following https://raw.githubusercontent.com/bytedance/deer-flow/main/Install.md

The agent automatically clones the repository, checks the environment and runs the initialization.

Option 2 – Manual Installation (≈5 minutes)

Prerequisites: git, Docker, Node.js 22+, uv, pnpm.

# 1. Clone repository
git clone https://github.com/bytedance/deer-flow.git
cd deer-flow

# 2. Run initialization wizard
make setup
# Choose model provider and enter API key

# 3. Check environment
make doctor

# 4. Start Docker
make docker-init
make docker-start

# 5. Open browser to begin using

Comparison with Mainstream Agents

Compared with Hermes, DeerFlow differs in several dimensions:

Developer : Hermes – US; DeerFlow – ByteDance (China).

Positioning : Hermes is a tool‑calling agent; DeerFlow is a “super‑agent” that handles end‑to‑end workflows.

Execution Mode : Hermes runs user‑provided tools; DeerFlow performs autonomous planning, sub‑agent collaboration and sandboxed delivery.

Use Cases : Hermes targets single‑task automation; DeerFlow targets complete project execution.

Deployment : Hermes – primarily local; DeerFlow – Docker or local.

License : Hermes – not fully open source; DeerFlow – MIT.

GitHub Stars : DeerFlow has over 72 000 stars and topped GitHub Trending on 28 Feb 2026.

The core distinction is that DeerFlow aims to “replace an entire team” by autonomously planning, collaborating, correcting and delivering complete results.

Security Tips

Although code runs inside a sandbox, the developers advise:

Never embed sensitive credentials directly in configuration files.

Configure network isolation for production deployments.

Regularly review official security notices and update to newer versions promptly.

Conclusion

DeerFlow exemplifies the shift of AI agents from simple tool invocation toward autonomous team collaboration. Its rapid adoption—over 72 000 stars in 72 hours—shows strong community interest in “AI‑driven full‑workflow replacement.” It is worth trying for security researchers, developers, or content creators.

Project URL: https://github.com/bytedance/deer-flow<br/>Official site: https://deerflow.tech