DeepSeek 发布 Super-agent 架构提案，用 Router 调度多个编码智能体
DeepSeek Proposes Super-Agent Architecture to Automate AI Coding Workflows
DeepSeek Harness 团队于 10 月 31 日发布技术提案《Exploring Super-agent Design: Towards a More Autonomous Agent System》，提出用中心 Router 把任务分发给并行执行的 Worker 智能体，以减少开发者逐个管理编码会话的负担。模拟实验中该架构的 95 分位任务完成时间为 15.7 分钟，单会话方案为 29.5 分钟；每次运行约 300 条消息、约 20 个任务请求，每种配置测试 5 次，且任务负载为合成数据、编码任务由耗时工具模拟，测的是调度与协调能力而非写码能力。DeepSeek 同时在 GitHub 以 MIT 许可开源研究原型 dsh-superagent-prototype，支持单个仓库并可选监听飞书消息，并说明该原型并非官方产品、与内部实现不同，还提示了提示词注入和凭证泄露等安全风险。
推荐理由
AI 评分 64
DeepSeek 公开了多智能体调度架构的设计与模拟实验数据，并附上 MIT 许可的原型代码，可供关注编码智能体编排的读者参考。
主题
正文
Chinese AI company DeepSeek on October 31 published a technical proposal for a ‘super-agent’ architecture designed to coordinate multiple artificial intelligence agents with limited human intervention. The proposal, titled ‘Exploring Super-agent Design: Towards a More Autonomous Agent System’, was published by the DeepSeek Harness team on its official WeChat account.
The article outlines an approach to reduce the time developers spend managing individual AI coding sessions. DeepSeek estimated that a person could typically oversee three to five active agent sessions simultaneously, making human attention a constraint on deploying multiple AI assistants.
The proposed architecture separates coordination from execution. A central agent, called a Router, receives messages and distributes work to independent agents, or Workers, which execute tasks in parallel and report their results. The Router handles user interactions while the Workers perform assigned tasks.
The system is also designed to respond to events such as messages and code updates without requiring explicit instructions for every task. DeepSeek described an example in which an agent detects a software bug reported in a group chat, attempts to reproduce and fix it, and submits a proposed code change for human review.
In simulated experiments, DeepSeek’s super-agent architecture recorded a 95th-percentile task completion time of 15.7 minutes, compared with 29.5 minutes for a single-session approach using a neutral prompt and a subagent tool.
The experiment involved about 300 messages per run, including around 20 work requests, with each configuration tested five times. DeepSeek said the workload was synthetic and coding tasks were simulated using a tool that consumed time. The tests measured scheduling and coordination capabilities, not the system’s ability to write or fix code.
The proposal comes as AI developers expand coding assistants. Anthropic’s Claude Code and OpenAI’s Codex are among the tools designed to handle increasingly complex software development tasks.
DeepSeek has also published an open-source research prototype, dsh-superagent-prototype, on GitHub under the MIT License. The prototype supports a single GitHub repository and can optionally listen to messages on Feishu, a workplace collaboration platform.
The company said the public prototype is not an official product and differs from its internal implementation. Some architectural ideas in the technical article remain future development directions and are not implemented in the released code.
DeepSeek also acknowledged security and privacy risks associated with autonomous agents connected to external services, including prompt injection and potential credential exposure. The company described the public code as a research prototype rather than production-ready software.