DeerFlow 2.0: ByteDance's super agent harness

What it is

DeerFlow (Deep Exploration and Efficient Research Flow) is ByteDance's open-source super agent harness. It defines itself plainly: a harness that orchestrates sub-agents, memory, and sandboxes to do almost anything, powered by extensible skills. It has more than 82,000 stars and reached number one on GitHub Trending after the 2.0 launch on 28 February 2026.

2.0 is a ground-up rewrite, and the README says it shares no code with v1. The original Deep Research framework is still maintained on the main-1.x branch and accepts contributions, but active development moved entirely to 2.0. That detail matters: a large share of what is written about DeerFlow online describes the 1.x deep-research pipeline, which is a different system from what is in the repo today.

From deep research to super agent harness

1.x was positioned as deep research: multi-agent retrieval, planning, and report generation. 2.0 lifts the positioning from "an application for one class of task" to "a runtime that carries any task". The README's own table of contents is evidence of that shift: Skills & Tools (including Claude Code integration), Session Goals, Manual Context Compaction, Sub-Agents, Sandbox & File System, Context Engineering, Long-Term Memory. None of those seven is research-specific - they are the infrastructure problems every long-horizon agent has to solve.

Two deserve singling out. Manual Context Compaction concedes something most harnesses avoid: when automatic compaction fires and what it discards is usually outside the user's control, and a bad compaction makes a long task lose a key premise halfway through. Promoting manual compaction to a first-class capability hands context management back to the operator. Session Goals give a long task a durable anchor, the same idea as the "goal" concept in this repo's own AGENTS.md.

Running and deploying it

Docker is the recommended deployment. Local development starts with git clone plus make setup: an interactive wizard walks you through choosing an LLM provider, optional web search, and execution and safety preferences - sandbox mode, bash access, file-write tools - then generates a minimal config.yaml and writes your keys to .env, which the project says takes about two minutes. make doctor verifies the setup at any time with actionable fix hints, and running make support-bundle before opening an issue produces summary and draft files.

There is also an install path built for the agent era: the README hands you a single prompt to give Claude Code, Codex, Cursor, or Windsurf, telling the coding agent to clone the repo if needed and bootstrap it per Install.md, prefer Docker when available, and stop with the exact next command plus any config the user still owes. Writing installation docs in a form an agent can consume is the right instinct for a harness project.

Observability has three tracing paths: LangSmith, Langfuse, and Monocle. IM channels are built in. Search and crawling integrate Jina, Browserless, and BytePlus's own InfoQuest - the web fetch from all three resolves relative links and image sources against the requested page URL, so the returned Markdown carries complete destinations. There is a per-model request_admission option for pacing requests to stay inside provider RPM limits, disabled by default.

Sister project and model guidance

LLM Space is the sister project the README names directly: a desktop tool for prototyping agent ideas, inspecting each harness step, replaying failures, and benchmarking performance. The README calls it the secret weapon behind DeerFlow. That thread is worth following, because the hard part of a harness is not getting it to run - it is localizing which step broke afterwards.

The project recommends Doubao-Seed-2.0-Code, DeepSeek v3.2, and Kimi 2.5 for running DeerFlow, with a Volcengine Coding Plan entry point (a separate volcengine link for mainland China). That is ecosystem steering and should be read as advice rather than a prerequisite: the harness itself does not lock the model.

Boundaries, and what you have to judge yourself

The README carries a Security Notice section titled "Improper Deployment May Introduce Security Risks", followed by recommendations. A harness that can run bash, write files, and speak to IM channels by default is a real exposure if it reaches the public internet on default configuration. Read that section before deploying and turn sandbox mode on.

The README itself runs to roughly 190,000 characters - high information density, but also a non-trivial onboarding cost. We have not benchmarked or reproduced it; the star count and the Trending number measure attention. It belongs at the top of the harness ladder because it is one of the largest and most thoroughly documented open-source specimens on the super-agent-harness route. Capability level is graded as needing reproduction.