FIND-20260323-008 · 2026-03-23 · Innovation Veille

DeerFlow by ByteDance — SuperAgent harness (37K+ stars)

trending-repo HIGH
Open-source SuperAgent harness from ByteDance that researches, codes, and creates. Handles tasks from minutes to hours using sandboxes, memories, tools, skills, and subagents. Built on LangChain/LangGraph. 37K+ stars, 4.4K+ forks. Supports multi-agent workflows, deep research, podcast generation. Python + TypeScript + Node.js.

Source

https://github.com/bytedance/deer-flow

ODS Impact

Highly relevant to our ADLC architecture: (1) their SuperAgent pattern with subagents mirrors our orchestrator+subagent model — compare delegation strategies, (2) sandboxes for code execution — we don't sandbox dev agents, potential safety improvement, (3) memories system — compare with our agent-memory/ persistence, (4) skills architecture — compare with our .claude/skills/ pattern, (5) multi-hour task handling — our agents timeout at context limits, study their approach to long-running tasks. ByteDance engineering quality at 37K stars makes this a serious reference.

Security Review

License: MIT | Maintenance: VERY ACTIVE (37K+ stars, ByteDance-backed, updated today) | Risk: LOW — MIT license, major corporate backing | Recommendation: Deep-dive architecture study. Compare agent delegation, sandbox, and memory patterns against ADLC.

Tags

multi-agent superagent bytedance langgraph sandboxes deep-research skills