Deer-flow

Deer-flow is ByteDance's open-source answer to long-horizon agentic tasks: problems that require sustained reasoning across many steps, tool calls, and sub-tasks before producing a final result. Where a simple coding agent handles one request in one session, Deer-flow is built for research workflows, end-to-end software builds, and complex information tasks that a single-session agent would time out on or lose track of.

The architecture combines sandboxed execution environments (so agents can run code safely without affecting the host system), a persistent memory layer, a library of tools and agent skills, a subagent dispatch system for breaking work into parallel or sequential chunks, and a message gateway that coordinates communication between agents. Together these allow the system to handle tasks that range from deep research synthesis to multi-file codebase changes with review loops.

Deer-flow sits in a category of long-horizon SuperAgent harnesses alongside tools like Manus and OpenClaw, but is notable for being a fully open-source release from ByteDance with active community development. Builders using it typically do so as a foundation to customize, not a hosted service to plug into. It is most relevant for teams building AI research assistants, internal knowledge workers, or complex automation pipelines where standard single-session agent frameworks run out of runway.