On the evening of October 31, the DeepSeek Harness team officially announced the underlying architecture for its exploration of autonomous agents—Super-agent (superintelligent agent)—and open-sourced it on GitHub.
This achievement is not a standalone new product. Instead, it introduces classic ideas from computer systems—such as scheduling, caching, storage-compute separation, and watchdog—into the agent architecture, attempting to take over session management and context synchronization at the underlying layer. In doing so, it aims to free up human attention and drive the agent application experience from a passive chat window toward an always-on autonomous collaborator.
Faced with the limitations of traditional single-session agents in multi-task concurrency, context confusion, and human intervention, the core architecture of Super-agent consists of five major modules. In its two-layer architecture, the system is split into a resident Router layer (responsible for unified routing and external interaction) and an isolated, concurrent Worker layer (focused on executing specific tasks), effectively balancing global continuity and high-concurrency throughput.
The event-driven module uses a two-stage filtering mechanism with fast screening at the program layer and slow judgment at the semantic layer. Combined with subscription and pass-through, Inline & Spill context organization, and background speculative execution, it enables the agent to proactively take semi-structured events from the environment as input and advance autonomously. In addition, asynchronous full-duplex interaction, a “brain-hand separation” design based on storage-compute separation, and a control loop formed by an external Watchdog together ensure the system’s persistent online availability and anomaly self-healing capabilities.
Preliminary experimental data show that the Super-agent based on this approach performs excellently in task completion efficiency: half of the tasks can be completed within 12.4 minutes, and 95% of tasks within 15.7 minutes, significantly outperforming traditional single-session solutions.
Meanwhile, the subscription pass-through mechanism saves the Router layer about 37% of Token overhead, while the batching mechanism reduces the amortized inference time per decision from 6.4 seconds to 1.6 seconds. Although the current experiments are based on a simplified environment and still face challenges such as two-layer scaling and security and privacy, this architecture preliminarily proves the feasibility of having the system layer take over session scheduling, providing an important industry reference for balancing efficiency and stability in multi-task agents.
▲ Super-agent open-source project address
GitHub address:
https://github.com/deepseek-ai/dsh-superagent-prototype
