bytedance/deer-flow
ByteDance's open-source super agent harness that spawns sub-agents, runs sandboxed code, and persists memory across sessions to automate complex multi-step tasks. Originally a deep research tool, version 2.0 became a full agent runtime — orchestrating parallel child agents, Docker-sandboxed execution, LangGraph workflows, and MCP server plugins from a single unified interface. Hit #1 GitHub Trending in February 2026 with nearly 30K stars.
Why It Matters
Most 'agent frameworks' still require you to manually chain prompts or write orchestration code. DeerFlow treats sub-agent spawning, sandboxed execution, persistent memory, and skill loading as first-class primitives — which means you describe the task, and a lead agent figures out which child agents to spin up, what code to run safely inside Docker, and how to synthesize the results. The Slack, Telegram, and Feishu integrations let non-engineers submit complex multi-step tasks from a chat window.