DeepSeek open-sources Super-agent research prototype
DeepSeek Harness published a technical article and open-sourced a Super-agent prototype under the MIT license on October 31. The Router–Worker system is designed to route work, filter messages, run tasks asynchronously, preserve state and monitor for stalls, and DeepSeek says it has used the system internally for bug reproduction, fixes and code review. In DeepSeek's simulation of about 300 messages and 20 work requests, Super-agent recorded a 15.7-minute P95 completion time, versus 29.5 minutes for a Session+Subagent setup using neutral prompts, but the open version remains a partial prototype and asynchronous execution is not enabled by default.
The release provides an open implementation for coordinating multiple coding agents beyond a single session, accompanied by vendor-run efficiency results, though real-world performance and scalability remain unverified.
一个程序员，同时指挥五个AI干活，会是什么体验？
一个负责修Bug，一个负责写代码，一个负责测试，另外两个在推进其他项目。看上去，五个AI同时开工，效率应该大幅提升。可现实往往没那么简单。
程序员需要在五个窗口之间来回切换，记住每个任务进行到哪一步、哪个需求发生了变化、哪个AI还缺少信息。忙活一圈，代码是不用自己写了，人却成了五个AI的项目经理。
AI越来越能干，人反而越来越忙。
10月31日晚，DeepSeek Harness团队发布技术文章《Super-agent设计探索：迈向更自主的Agent系统》，并同步开源Super-agent技术原型，试图解决这一难题。
与过去接到指令后执行任务的Agent不同，Super-agent希望让AI自己分配工作、主动发现问题，并持续跟进任务，不再事事等待程序员安排。