Deer-Flow

DeerFlow is an open-source AI agent framework developed by ByteDance and hosted on GitHub (bytedance/deer-flow), built on LangGraph and LangChain, for creating agents that carry out extended, multi-hour research, coding, and content-creation tasks rather than single-turn responses. A lead agent can orchestrate specialized sub-agents with their own isolated context, running them in parallel and synthesizing their results back into a structured output. Each agent runs inside a sandboxed execution environment, either locally, in Docker/AIO containers, or provisioned on Kubernetes, so it can execute bash commands, manipulate files, and browse the web via an optional Playwright-based browser tool, without exposing the host system. Capabilities are loaded progressively through a Markdown-based skills system rather than all at once, and users can enable, disable, or write custom skills, or connect third-party tools through MCP servers. DeerFlow also maintains persistent, cross-session user memory (DeerMem, with optional mem0/OpenViking integration) and supports tracing through LangSmith, Langfuse, and Monocle for observability. The framework can receive tasks through messaging platforms including Telegram, Slack, Feishu, WeChat, WeCom, and DingTalk without requiring a public IP, and its Gateway API handles checkpointing (SQLite/Postgres) and multi-worker coordination via Redis. It's intended for developers and enterprises building production-grade, long-running agents that need auditability and transparent execution, and has attracted a large open-source following, with roughly 79,000 GitHub stars.