DeepSeek is open source again! Announcement of Super agent architecture

Zhidongzhi author | Li Shuiqing Editor | ▲Super-agent open source project address GitHub address: https://github.com/DeepSeek-ai/dsh-superagent-prototype The “Super-agent” that DeepSeek wants to do can be generally understood as an Agent that is always online, can do multiple things at the same time, can take the initiative to find work, and can recover by itself when problems occur. The result launched by the team this time is not a new product, but closer to a new idea - moving classic ideas such as scheduling, caching, separation of storage and calculation, and watchdog (Watchdog) in computer systems into the Agent architecture. As the model performs single tasks with less human intervention, the bottleneck shifts to human attention. DeepSeek Harness attempts to take over session management and context synchronization at the bottom through five major modules - dual-layer architecture, event-driven, asynchronous multi-channel interaction, storage and calculation separation and external Watchdog, allowing people to move up to decision-making and acceptance, and improve the Agent application experience. ▲Super-agent overall system architecture Experimentally, half of the Super-agent tasks based on this idea were completed in 12.4 minutes, and 95% of the tasks were completed in 15.7 minutes It is completed within one session, which is significantly better than a single Session with the Subagent tool; after removing the subscription pass-through, the number of Router decisions increases by about 30%, and the read cache token increases from 23.23 million to 31.77 million; Router batch processing can reduce the inference time of a single decision from 6.4 seconds to 1.6 seconds. This set of data means that Super-agent has proven the feasibility of "system layer taking over session scheduling" in preliminary experiments.