DeerFlow

DeerFlow is an open-source "SuperAgent harness" developed by ByteDance. It is a high-level orchestration framework designed to transform Large Language Models (LLMs) into autonomous agents capable of completing complex, "long-horizon" tasks that typically take minutes to hours—such as deep research, coding, and multi-step content creation.

What It Is

DeerFlow (Deep Exploration and Efficient Research Flow) is an open-source, full-stack AI agent system built on LangGraph and LangChain. Unlike traditional chatbots that suggest actions, DeerFlow is an execution-first platform. It provides a "computer" for AI agents, allowing them to interact with a real filesystem, execute code in secure sandboxes, and manage long-running workflows without human intervention. Originally an internal tool at ByteDance, it was open-sourced to provide a production-ready infrastructure for autonomous agents.

What It Can Do

DeerFlow enables AI agents to move beyond text generation into active task execution:

• Autonomous Execution: Runs code (Python/Bash) and manages files directly within isolated environments.
• Multi-Agent Orchestration: Decomposes a single complex prompt into multiple sub-tasks and spawns specialized sub-agents to handle them in parallel.
• Stateful Memory: Maintains both short-term thread context and long-term user memory (preferences, styles, and past facts) across different sessions.
• Deep Web Research: Searches the web, crawls pages, and synthesizes information from diverse sources into structured documents.
• Extensible Skills: Uses a modular "Skill" system (defined via Markdown files) that tells the agent how to handle specific domains like "Frontend Design" or "Data Analysis."

Examples of Its Capabilities

• Deep Research Reports: Given a topic like "2026 AI Trends," it can search dozens of sources, extract data, generate charts via Python, and output a formatted report with citations.
• Full-Stack Scaffolding: It can write the code for a web application, set up the directory structure, run tests in its sandbox, iterate on errors, and deliver a working project.
• Multimedia Content: It can generate PowerPoint slide decks (using Marp), create audio podcasts from research summaries, and even generate images/videos based on literary descriptions.
• Data Pipelines: It can ingest a raw CSV, write and execute a data-cleaning script, and produce a visualized exploratory data analysis (EDA).

How Does It Work?

DeerFlow operates through a layered architecture:

1. Lead Agent (Supervisor): Acts as the brain. It analyzes the user's request, selects necessary "Skills," and creates a plan.
2. Sandbox Execution (AIO Sandbox): Every task runs in an isolated Docker container (or Kubernetes pod). This environment contains a full filesystem, a web browser, and a shell terminal, ensuring the agent's actions are safe and persistent.
3. Sub-Agent Spawning: For heavy tasks, the Lead Agent spawns parallel sub-agents (e.g., a "Researcher" to find data and a "Coder" to process it).
4. Context Engineering: To prevent "context window" overflow during long tasks, DeerFlow summarizes completed sub-tasks and offloads intermediate data to the filesystem, keeping the active reasoning window focused.
5. Human-in-the-Loop: It allows users to review plans or intermediate results before the agent proceeds to the next phase.

Applications of DeerFlow

The applications of DeerFlow span several high-impact domains, ranging from enterprise research—where it automates competitive analysis, market research, and technical due diligence—to software development, facilitating autonomous bug fixing, documentation generation, and unit test automation. It is equally effective in content marketing for generating multi-format assets like slide decks and social media posts from a single research brief, as well as in data science for automating complex report generation and building real-time dashboards directly from raw datasets.

Previous Models 

• DeerFlow 1.0 (Launched May 2025): The original version was a specialized framework focused primarily on "Deep Research." It was more limited in scope, focusing on gathering information and summarizing it.
• DeerFlow 2.0 (Released Feb 2026): A ground-up rewrite that transitioned the project from a research tool to a "Super Agent Harness." Version 2.0 introduced the full-stack architecture, persistent memory, and the "execution-first" sandbox model that allows for coding and system-level tasks.
• Internal Prototype: Before the open-source release, DeerFlow existed as an internal ByteDance tool for automating complex internal documentation and research workflows.

Note: DeerFlow is model-agnostic but is optimized for high-reasoning models with long context windows (100k+ tokens), such as Doubao-Seed-2.0, DeepSeek v3, and Claude 3.5/3.7.