Anthropic and DeepSeek propose compatible agent architectures
A commentary argues that Anthropic’s recommendation to fold subagents into a single agent loop and DeepSeek Harness’s persistent Router/Worker design address different architectural layers rather than contradict each other. Anthropic emphasizes a single loop with skills and tools for capabilities that stronger models can handle internally, while DeepSeek proposes a persistent Router coordinating isolated Workers across concurrent tasks and sessions. The commentary also cites performance figures comparing DeepSeek’s Super-agent design with a single-session subagent approach; these figures come from the cited material and were not independently verified in the supplied evidence.
The analysis provides a practical framework for agent-system design: simplify orchestration within individual tasks, while adding persistent coordination when workloads span multiple concurrent tasks or sessions.
原文地址:
本文并非原文的翻译,是经过我自己理解后的输出,这是最近关注最多的两篇 Agent 的文章了。我挑选了自己感兴趣的部分,让 AI 辅助理解并总结。感兴趣的朋友也可以查看原文。
表面上,两家在"打架"
最近这段时间,两家公司先后抛出了两份关于 Agent 架构的论述,方向看起来完全相反。
Anthropic 在《When to fold subagents into your main agent loop》里说:企业把系统拆成一堆 subagent 的做法,大多是在替模型做它已经不再需要的事。正确的方向是把 subagent 折叠回单个 agent loop,用 skills 和 tools 来承载领域知识。
DeepSeek Harness 团队在《Super-agent 设计探索》里却说:要打破 Agent 与单次会话的绑定,引入一个常驻的 Router 层,下面挂一群隔离的 Worker 去并发干活。
#Anthropic#DeepSeek#DeepSeek Harness#Super-agent#Claude Code#Opus 4.5