bytedance/deer-flow
An open-source long-horizon SuperAgent harness that researches, codes, and creates. With the help of sandboxes, memories, tools, skill, subagents and message gateway, it handles different levels of tasks that could take minutes to hours.
DeerFlow is ByteDance's open-source framework for long-horizon AI agent tasks that take minutes to hours rather than a single turn. Built on LangGraph, it orchestrates subagents for deep research, code writing, and content creation with sandboxed execution and persistent memory in 2026.
Our Review
ByteDance open-sourced DeerFlow in early 2025 as a response to a category of AI work that standard chatbots cannot handle: research-grade tasks that require browsing dozens of sources, synthesizing findings, writing code to test conclusions, and producing structured outputs -- all in one unattended run. The ByteDance team positioned it as a long-horizon agent system and built it on LangGraph to manage the multi-agent state machine that makes hour-scale workflows tractable. With 68,000 GitHub stars, the project attracted teams who needed autonomous research and creation beyond what single-turn agents could do.
Key capabilities
- •Long-horizon task execution: designed for workflows that take minutes to hours, not single-turn completions -- research reports, multi-step code projects, podcast episodes
- •Multi-agent orchestration: a coordinator dispatches specialized subagents for research, code writing, content creation, and quality review, then aggregates their outputs
- •Sandboxed code execution: code written by the agent runs in an isolated environment, preventing unsafe operations from reaching the host system
- •Persistent memory: agents retain context across multi-step tasks and across sessions, enabling resumable long-running workflows
- •Tool integration: web search, URL scraping, file operations, and code execution are available as native tools for any subagent
- •Message gateway: structured communication protocol between subagents prevents context collapse on complex multi-agent runs
Getting started
Clone bytedance/deer-flow, install dependencies with pip install -r requirements.txt, and configure your LLM API key in the .env file. Launch the server with python -m deerflow.server, then submit a research task via the REST API or the included web interface. DeerFlow handles the rest of the workflow automatically.
Limitation
Long-horizon runs on complex tasks can consume significant LLM API tokens -- a deep research report may cost $5-20 in API calls depending on model and task scope. The LangGraph dependency means debugging multi-agent state issues requires familiarity with LangGraph's execution model. Token limits on individual LLM calls can constrain the agent's working memory on very large documents or codebases. No built-in billing or rate-limit management for the underlying LLM API.
Our Verdict
DeerFlow fills a gap that standard AI assistants leave open: tasks that are genuinely multi-step, require synthesizing heterogeneous sources, and produce structured outputs on the order of hours of work. The ByteDance engineering background shows in the LangGraph-based orchestration and the message gateway design -- both choices reflect production-grade thinking about state management across subagents, not prototype-grade sequential tool calls.
The long-horizon positioning is the honest differentiator. Where most agent frameworks compete on single-task quality, DeerFlow's architecture is designed around the assumption that the task will require many sequential decisions, some of which depend on outputs from earlier steps. The memory system and message gateway exist because that assumption drives the design.
The practical constraint is cost. A deep research run that processes 20+ sources and produces a structured report can consume meaningful LLM API budget. Teams evaluating DeerFlow should budget for API costs alongside infrastructure and test with smaller task scopes before scaling to production workloads in 2026.
Frequently Asked Questions
What is DeerFlow and how does it differ from a standard AI assistant?
DeerFlow is ByteDance's open-source multi-agent framework designed for long-horizon tasks -- work that takes minutes to hours and requires coordinating multiple AI agents. A standard assistant handles one turn at a time. DeerFlow manages a workflow: a coordinator dispatches specialized subagents for research, coding, and content creation, tracks state across all of them, and produces a final aggregated output. It is built for unattended operation on complex tasks in 2026, not interactive conversation.
What kinds of tasks is DeerFlow designed to handle?
DeerFlow is designed for research reports, multi-step code projects, podcast episode generation, and content creation tasks that require synthesizing multiple sources. Concretely: give it a research question and it browses sources, synthesizes findings, writes a structured report, and optionally produces a podcast episode summarizing the research. It handles tasks where a human would normally spend 30 minutes to several hours doing sequential research and writing in 2026.
Which LLMs work with DeerFlow?
DeerFlow supports any OpenAI-compatible model as its backend. In practice, users run it with Claude Sonnet, GPT-4o, Gemini 2.5 Pro, and DeepSeek via OpenRouter. The quality of the agent's research and synthesis scales with model tier -- frontier models produce substantially better structured outputs on complex multi-step tasks. Smaller models can be used to reduce API costs but produce less reliable results for deep research workflows in 2026.
How much does a DeerFlow research run cost in API tokens?
Cost depends on task complexity and model choice. A simple research task on a narrow topic with 5-10 sources typically costs $1-3 with a frontier model. A deep research report requiring 20+ sources, cross-referencing, and structured output can cost $10-20 per run. The multi-agent architecture means each subagent incurs its own LLM costs. Setting token budgets per run and testing with cheaper models first is the recommended approach before running expensive production workloads in 2026.
How does DeerFlow compare to OpenHands or other autonomous coding agents?
OpenHands is optimized for software development tasks -- writing code, fixing bugs, opening pull requests. DeerFlow is optimized for research-first workflows where the output may be a report, a podcast, or a document rather than a code commit. DeerFlow can write code as part of a larger workflow, but its core use case is long-horizon information synthesis rather than software engineering. The two tools are complementary for teams doing both in 2026.
What is deer-flow?
DeerFlow is ByteDance's open-source framework for long-horizon AI agent tasks that take minutes to hours rather than a single turn. Built on LangGraph, it orchestrates subagents for deep research, code writing, and content creation with sandboxed execution and persistent memory in 2026.
How do I install deer-flow?
Visit the GitHub repository at https://github.com/bytedance/deer-flow for installation instructions.
What license does deer-flow use?
deer-flow uses the MIT license.
What are alternatives to deer-flow?
Explore related tools and alternatives on My AI Guide.
Open source & community-verified
MIT licensed: free to use in any project, no strings attached. 68,829 developers have starred this, meaning the community has reviewed and trusted it.
Reviewed by My AI Guide for relevance, quality, and active maintenance before listing.
Topics