DeerFlow began as a deep-research pipeline; version 2 threw that code away to chase a harder question — what does an agent actually need to run unsupervised for hours instead of seconds? Its answer is infrastructure over prompt tricks: durable memory, sandboxed execution, and a lead agent that fans work out to isolated sub-agents rather than looping one model on itself.
What Sets It Apart
- Skills are the extension unit, not one-off tools. Behavior is packaged as reusable skills the harness composes, so adding a capability doesn't mean rewriting an agent loop.
- Sub-agents fan out, they don't chat. A lead agent spawns sub-agents on the fly, each with its own scoped context, tools, and stop conditions; they run in parallel and report structured results back for the lead to synthesize. Every sub-agent is context-isolated — it can't see the main agent or its siblings, which keeps long jobs from drowning in each other's tokens.
- Sandboxes with a real filesystem. Code runs isolated but with persistent files, so a multi-hour job survives across steps instead of living only inside a prompt.
- The gateway faces clients, not the agents. An HTTP, LangGraph-compatible gateway fronts the runtime so external clients, IM channels, and skill installs reach it; the "messages" here are chat events and SSE streams, not sub-agent traffic.
- Model-agnostic on LangChain/LangGraph. Swap Doubao, DeepSeek, Kimi, GPT, or Claude through config rather than code.
Who It's For
Great fit if you want an open-source harness for long-horizon, multi-step automation — research, coding, and content generation that runs for minutes to hours — and you're comfortable wiring up sandboxes and model keys yourself. Look elsewhere if you need a lightweight chat UI or a plug-and-play SaaS: this is self-hosted infrastructure, and the docs are blunt that careless sandbox or deployment settings carry real security risk.
