Reportedly: DeepSeek Harness open-sources Super-agent architecture prototype
Reported by Zhidongxi and QbitAI · not yet confirmed by the company or a second independent outlet. We update this page when it is.
DeepSeek Harness released the dsh-superagent-prototype on GitHub, a Super-agent architecture that keeps context across sessions and runs multiple tasks in parallel. The prototype uses a two-layer Router/Worker design, event-driven async channels, compute-storage separation, and an external Watchdog. In tests, 50% of tasks finished within 12.4 minutes and 95% within 15.7 minutes, Zhidongxi reported.
China context
- Original name
- 深度求索
- Outside China
- Open weights · zhidx.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- The open-sourced prototype provides a concrete reference for building persistent, multi-task agents; developers can inspect the Router/Worker separation and Watchdog patterns.
- For investors
- The release signals DeepSeek's direction toward autonomous agent infrastructure, but no commercial product or pricing is announced, so revenue impact is unproven.
Translated from Chinese. Quotes and facts link to the original sources.
DeepSeek Harness published a technical article and open-sourced a prototype for Super-agent, an agent architecture designed to operate continuously, handle multiple tasks concurrently, and recover from failures. The design applies classic system concepts—scheduling, caching, compute-storage separation, and Watchdog—to agent systems. The team reports that internal instances already run continuously, participating in issue reproduction, code review, and DeepSeek Harness iteration. Experimental results show half of tasks completed in 12.4 minutes and 95% in 15.7 minutes, outperforming a single-session agent with subagent tools. Removing subscription pass-through increased Router decisions by about 30% and read cache tokens from 23.23 million to 31.77 million; Router batching reduced per-decision inference time from 6.4 seconds to 1.6 seconds.
The architecture separates a Router layer from Worker layer, uses event-driven asynchronous multi-channel interaction, and external Watchdog for recovery. The reported trade-off is higher token consumption for faster, more stable multi-task execution. The experiments used synthetic workloads and simplified environments, so results are preliminary.
Developers building persistent personal assistants get a reference architecture that removes the need to manually switch sessions and maintain context, potentially reducing the human attention bottleneck in multi-task agent workflows. The open-sourced prototype lets outside-China teams inspect and adapt the design, but no production-ready product or API is announced.
For teams building always-on agents, the architecture offers a blueprint for session management, context synchronization, and failure recovery. The open-source release lowers the cost of experimentation, but the lack of a packaged product means integration effort remains.
Next verifiable signals: whether DeepSeek releases a production Super-agent service or API, whether the prototype is adopted in third-party agent frameworks, and whether independent benchmarks confirm the reported task completion times.