DeepSeek open-sources the dsh-superagen multi-agent development project

Liang Sheng is awesome—everyone stand up. I highly recommend this 14,000-character technical long-read on Super-agent that DeepSeek just published. Damn, it finally rips off Agent's biggest fig leaf: When things go badly, it's often not that the model isn't strong enough, but that people simply can't manage it all. It wouldn't even be an exaggeration to say it overturns the old dialog-box-centric paradigm. It does just one core thing: completely frees people from the drudgery of managing conversations. Many people think the bottleneck to Agent deployment lies in model capability. In the article, DeepSeek points out the truth: the real bottleneck is human attention. In other words, once single-task execution no longer needs a human watching it, the old product form built around a single dialog box should be eliminated. Everyone assumes by default that the biggest obstacle to deploying large models is insufficient reasoning capability, and the whole internet is desperately grinding on parameters and chasing benchmark scores, trying to make a single dialog box handle everything. But the real dividing line has never been at single-point intelligence; it lies in the system's ultimate release of human attention. A normal person can follow at most 3 to 5 active conversations at the same time. If the system always dumps the burden of switching windows, chasing progress, and sorting out context onto the user, then no matter how much compute it has, it's just a toy that exhausts people to death. DeepSeek's solution is extremely cold-eyed: completely shatter the single-conversation mechanism and build a long-lived Super-agent architecture. The front end keeps a lightweight Router that only handles dispatching work from multiple channels; in the middle, Workers with fully isolated contexts are spawned to focus on single-point tasks; follow-up events for old tasks are subscribed to and delivered directly to the person in charge; and an independent Watchdog is added to patrol for infinite loops. Task completion time drops directly from 29.5 minutes with a single conversation to 12.4 minutes, and replies in chat software are even hard-capped at under 100 characters—never disturbing anyone when it shouldn't speak. This is an extremely sobering reminder for all teams doing AI deployment and business management: the core moat of the future is not how well the model answers questions, but who can reduce chaotic conversation scheduling to a self-healing pipeline that needs no human babysitting.

View original

梁圣牛逼，全体起立，DeepSeek 刚发布的这篇1.4 万字的Super-agent 技术长文强烈推荐，喵的终于把 Agent 最大的遮羞布扯下来了： 做不好往往不是模型不够强，而是人根本管不过来。 甚至说推翻了以对话框为中心的旧范式都不为过，核心就做了一件事：把人从管理会话的苦力活里彻底解脱出来。 很多人以为 Agent 落地的瓶颈在模型能力，DeepSeek 在文章里指出了真相：真正的瓶颈在人的注意力。 也就是说当单任务执行不再需要人盯，以一个对话框为基本单位的旧产品形态该被淘汰了。 大家默认以为大模型落地的最大阻碍是推理能力不够强，全网都在拼命卷参数、刷跑分，试图靠单个对话框包办一切。 但真正的分水岭从来不在单点智力，而在系统对人类注意力的极致释放。 一个正常人同时最多只能跟进 3 到 5 个活跃会话，如果系统永远把切窗口、催进度、理上下文的负担甩给用户，算力再高也只是个把人累死的玩具。 DeepSeek 给出的解法极其冷峻：彻底打碎单会话机制，搭一套长期存活的 Super-agent 架构。 前端常驻一个轻量 Router 只管对接多渠道派活，中间派生出上下文完全隔离的 Worker 专攻单点任务，老任务后续事件靠订阅直通直达负责人，再配上一个独立的 Watchdog 巡检死循环。 任务完成时间直接从单会话的 29.5 分钟压到 12.4 分钟，聊天软件里的回复甚至被焊死在 100 字以内，不该发声时绝不打扰。 这对所有做 AI 落地和业务管理的团队是一个极度清醒的提醒：未来的核心壁垒不在于模型多会答题，而在于谁能把混乱的会话调度降维成一条不用人肉盯梢的自愈流水线。