Complete Guide to DeerFlow 2.0: ByteDance's Open-Source SuperAgent Framework That Hit #1 on GitHub

On February 27, 2026, ByteDance quietly open-sourced a project called DeerFlow 2.0 — short for Deep Exploration and Efficient Research Flow — and within 24 hours it was sitting at the top of GitHub Trending with roughly 25,000 stars and 3,000 forks. The attention was not accidental. DeerFlow 2.0 is a complete ground-up rewrite of the original DeerFlow framework, sharing no code with version one, and it represents something genuinely different from the growing roster of AI agent frameworks like CrewAI, AutoGen, and LangGraph. Where most agent frameworks give you components to assemble, DeerFlow ships as a fully opinionated, batteries-included system — a "SuperAgent harness" that orchestrates sub-agents, persistent memory, sandboxed execution environments, and an extensible skills layer to autonomously handle complex, multi-hour tasks ranging from deep web research to full application prototyping.

The architecture at the heart of DeerFlow 2.0 is built on LangGraph and LangChain, but what sets it apart is the execution-first philosophy. Rather than agents producing text suggestions or code snippets that a human must copy and run, DeerFlow gives each agent an actual computer: an isolated Docker container with a full filesystem, a bash terminal, and the ability to read, write, and execute files in real time. When you hand DeerFlow a complex prompt, a lead agent acts as the orchestrator, decomposing the request into structured sub-tasks, deciding which tasks can run in parallel, and spawning specialized sub-agents to handle each one. A research task might fan out into a dozen sub-agents simultaneously — each exploring a different angle, querying different sources, or generating different assets — and then converge the results into a single coherent output. That output might be a detailed research report with cited sources, a complete slide deck with generated visuals, a functional web page, or even an AI-generated video with reference images. The key insight is that DeerFlow does not just plan work; it executes it end-to-end within secure sandboxes, which makes it meaningfully closer to an autonomous software engineering assistant than a chatbot with tool access.

The skills system deserves particular attention because it is how DeerFlow achieves extensibility without losing its opinionated defaults. Skills are defined as Markdown files that encode workflows and best practices — think of them as reusable playbooks that the agent can load progressively as context demands. Out of the box, DeerFlow ships with built-in skills for web research (gathering cited sources across the entire internet), report generation, slide deck creation, web page generation, and media creation. But any team can write custom skills that encode domain-specific workflows, and the agent will load them as needed rather than stuffing everything into the initial prompt. This progressive loading approach is a practical response to the context window management problem that plagues most agent systems — rather than front-loading every possible instruction, DeerFlow pulls in only what is relevant to the current sub-task.

Memory is another area where DeerFlow 2.0 makes an opinionated bet. The framework includes a persistent memory layer that tracks user preferences and project context across sessions, using asynchronous debounced queues to avoid blocking the agent's primary execution loop. ByteDance recently added TIAMAT as a cloud memory backend option for teams that need shared state across multiple users or deployments. That said, even ByteDance acknowledges that agent memory remains an unsolved problem in practice — the memory layer works well for preference tracking and session continuity, but true long-term reasoning over accumulated knowledge is still an open research question across the entire field.

For developers evaluating DeerFlow 2.0, the model-agnostic architecture is a significant draw. The framework integrates with any OpenAI-compatible API endpoint, which means you can run it against GPT-4, Claude, Google Gemini, DeepSeek, Kimi, or local models served through Ollama — all through the same configuration interface. ByteDance's documentation specifically recommends Doubao-Seed-2.0-Code, DeepSeek v3.2, and Kimi 2.5 for optimal results, noting that strong instruction-following capabilities matter most for reliable task decomposition. For teams already invested in Google's Gemini ecosystem, DeerFlow supports Gemini models through OpenRouter, making it straightforward to build autonomous agent workflows powered by Gemini 2.5 Pro or the newly released Gemini 3.1 while keeping full control over orchestration, memory, and execution locally.

Deployment flexibility is another practical advantage. DeerFlow offers three modes: local development (for quick iteration), Docker (for single-server production), and Kubernetes (for enterprise-scale multi-server deployments). The standard architecture runs four processes — a Gateway API, a LangGraph Platform server, and supporting services — while an experimental Gateway Mode collapses this into three processes by embedding the agent runtime directly in the Gateway API for faster cold starts and lower resource consumption. The entire project is released under the MIT License, meaning any organization can use, modify, and deploy it commercially with no restrictions, no usage caps, and no attribution requirements. For enterprises evaluating open-source agent frameworks, this permissive licensing combined with Docker and Kubernetes support makes DeerFlow one of the most deployment-ready options available today.

The framework also ships with built-in observability through LangSmith and Langfuse tracing integrations, plus native messaging channel support for Telegram, Slack, Feishu/Lark, and WeCom — all configured through long-polling or WebSocket connections that require no public IP address. This means teams can deploy DeerFlow behind a corporate firewall and still get real-time agent status updates in their existing communication tools. The combination of sandboxed execution, persistent memory, extensible skills, multi-model support, and enterprise deployment options makes DeerFlow 2.0 arguably the most complete open-source agent harness available as of early 2026 — though teams should carefully evaluate the security implications of giving an AI agent genuine code execution capabilities, and ByteDance's country of origin may trigger additional review processes for some organizations.

Read the full project documentation at GitHub (bytedance/deer-flow) →