DeerFlow — SuperAgent for Research, Code & Creative

What it is

DeerFlow is an open-source SuperAgent framework from ByteDance designed for long-horizon tasks. It orchestrates sub-agents, sandboxed execution environments, persistent memory, and reusable skills to handle complex workflows that span minutes to hours. The framework covers research, code generation, and creative tasks with a web UI for monitoring agent progress.

DeerFlow targets developers and teams building autonomous agents that need to execute multi-step plans with coordination between specialized sub-agents. It provides the infrastructure for long-running agentic workflows with built-in error recovery and state persistence.

How it saves time or tokens

DeerFlow handles the orchestration layer that you would otherwise build manually. Sub-agent delegation means each agent focuses on its specialization (research, coding, writing) rather than one monolithic agent trying to do everything. Sandboxed execution provides safe code running without risking your system. Persistent memory across sessions means agents do not restart from scratch.

How to use

1. Clone the repository and set up configuration: git clone https://github.com/bytedance/deer-flow.git && cd deer-flow && make config.
2. Set API keys in .env for your LLM provider (OpenAI, Anthropic) and tools (Tavily for search).
3. Launch with Docker: make docker-start and access the UI at http://localhost:2000.

Example

# Clone and configure
git clone https://github.com/bytedance/deer-flow.git
cd deer-flow
make config

# Set API keys in .env
# OPENAI_API_KEY=sk-...
# TAVILY_API_KEY=tvly-...

# Launch via Docker
make docker-start

# Access the UI at http://localhost:2000
# Submit a task: 'Research the latest advances in protein folding
# and write a summary report with citations'

Related on TokRepo

• Agent Tools — AI agent frameworks and orchestration tools
• Research AI Tools — AI-powered research and analysis tools

Common pitfalls

• DeerFlow requires multiple API keys (LLM, search, etc.) to function fully. Missing keys will silently disable capabilities rather than failing with clear errors.
• Long-horizon tasks consume significant tokens. Monitor costs, especially when sub-agents iterate on research or code generation.
• The Docker deployment is recommended over local setup due to sandboxing requirements. Local execution without containers may expose your system to code execution risks.