DeerFlow (Deep Exploration and Efficient Research Flow) is ByteDance's open-source super agent harness for building autonomous systems that research, code, and create. Released under the MIT license, it orchestrates sub-agents, memory, sandboxes, and extensible skills to handle tasks that range from a few minutes to several hours. With over 73,000 GitHub stars, it became one of the most prominent agent frameworks of 2026, claiming the number one spot on GitHub Trending after the launch of version 2. ## A Super Agent Harness, Not Just a Chain DeerFlow positions itself as a harness rather than a single agent. Instead of running one model in a fixed loop, it coordinates multiple sub-agents that each take on part of a long-horizon task, sharing context through a message gateway and persistent memory. This design lets the system decompose complex objectives, delegate sub-tasks, and assemble results, which is what allows it to take on multi-step jobs that would overwhelm a single prompt-response cycle. ## Ground-Up 2.0 Rewrite DeerFlow 2.0 is a complete rewrite that shares no code with the original 1.x Deep Research framework, which remains maintained on a separate branch. The new version broadens the project's scope from deep research alone to a general-purpose agent platform, while keeping the research workflow that first made it popular. The rewrite reflects how quickly agentic tooling has matured, moving from single-purpose research bots toward configurable, production-oriented harnesses. ## Memory, Sandboxes, and Skills Three capabilities define what DeerFlow can do. Persistent memory lets agents retain and reuse information across steps; sandboxes provide isolated environments where the system can safely run code and execute tools; and an extensible skills system lets developers add new capabilities without rewriting the core. Together these give DeerFlow the ability to not only plan but actually act, by writing and running code, browsing, and producing artifacts like reports and podcasts. ## Built on a Modern, Integrable Stack The project is built around LangChain and LangGraph and spans a Python backend with a Node.js and TypeScript frontend. It supports Model Context Protocol servers for connecting external tools, integrations with instant-messaging channels, and observability through LangSmith and Langfuse tracing. A Docker-based deployment path is recommended, and the project documents deployment sizing so teams can match resources to workload. ## Practical Research and Creation DeerFlow is aimed at real deliverables, not demos. It can carry out deep research across the web, synthesize findings, generate code, and create multimedia output, and it integrates ByteDance's InfoQuest search-and-crawl toolset for retrieval. Recommended pairings with strong reasoning and coding models position it as a practical foundation for analysts, developers, and content teams who need an agent that follows a task through to a finished result. ## Considerations As a full harness, DeerFlow carries more operational complexity than a lightweight library: standing it up involves configuring models, sandboxes, and optionally MCP servers and tracing backends. Long-horizon autonomous runs that execute code and call external tools also demand careful sandboxing and cost monitoring, since multi-agent workflows can consume significant tokens and compute. Teams adopting it should plan for that setup and governance overhead alongside the productivity gains.