DeepSeek 甩出 Super-agent：AI 从"等人派活"变成"自己找活干" DeepSeek Harness 团队在公众号发布技术长文《Super-agent 设计探索：迈向更自主的 Agent 系统》，并开源原型 dsh-su

DeepSeek throws out Super-agent: AI changes from "waiting for others to assign jobs" to "finding jobs by itself" DeepSeek The Harness team released a long technical article "Super-agent Design Exploration: Toward a More Autonomous Agent System" on the public account, and open sourced the prototype dsh-superagent-prototype (MIT protocol). Officials point out that this is an engineering research prototype, not a ready-made product. But the scene in the article is very catchy - the internal running instance participates in defect recurrence and code review, and even if there is no one in the group, it can sniff out potential bugs and background recurrence. It is people, not models, that are stuck in multitasking. Up to now, mainstream Agents are based on a single Session, and sessions are separated from each other. If people want to do several things in parallel, they can only repeatedly switch windows and remember the context. The team observed that a person can usually only keep up with 3 to 5 active sessions, so a strange circle appears: the single-task limit depends on the model, but the multi-task bottleneck is stuck in the person's attention. It is this barrier that Super-agent wants to break. A dispatch center, a group of "hands" working at the core is a two-layer division of labor: the Router resident at the front desk manages interactions and assigns tasks, and a lightweight model can be used to support it; each task has a context-isolated Worker, which runs in parallel using the strongest model, and completes the return to the Router. Coupled with event-driven, storage-calculation separation and Watchdog supervision, stuck tasks will be automatically detected and repaired. The stress test included about 300 messages and 20 requests, with P50 taking 12.4 minutes, P95 taking 15.7 minutes, and the single-session solution taking P95 up to 29.5 minutes. The cost is even more severe - V4-Pro is about US$1.32/3.96 per million tokens, which is one order of magnitude lower than GPT-5.5 and Opus 5.5. When the upper limit of concurrency changes from "human attention" to "system throughput", the agent can be considered out of the chat window. via AI News (author: AI Base)