Three agent frameworks, 205k stars combined, converge on one shape

ByteDance rebuilt deer-flow from the ground up, and the result lands on the same five-part shape already running in LobeHub and HKUDS/nanobot. Three separate teams, 205,000 combined GitHub stars, all landed on identical architecture. That's a strong signal that the hard design questions in agent runtimes already have answers.

deer-flow's v2 runs sandboxes in three flavors: local, Docker, or Kubernetes. Each task gets its own full filesystem view. Long-term memory is file-backed, searched with SQLite FTS5, with per-agent facts kept in Markdown. Skills are also Markdown, and they load progressively instead of all at once. Sub-agents run in isolated contexts with scoped tools, and report back structured results. A message gateway ties it to Telegram, Slack, Feishu, WeChat, and DingTalk.

LobeHub, at 80,900 stars, and HKUDS/nanobot, at 46,300 stars, expose close to the same primitives. Three codebases, one shape. All three wire agents into consumer chat apps instead of a dedicated interface. That's where these teams expect agents to actually get used.

Skip this decomposition when you build your own runtime, and you're re-solving sandbox isolation and skill-loading that three teams already shipped. Watch whether any of them breaks from it in their next major version. That would be the tell the shape isn't settled after all.