DeepSeek’s Super-agent architecture already powers continuously running internal instances handling bug reproduction, fixes, code-review responses, and development work.

DeepSeek’s Harness team published “Super-agent Design Exploration: Toward More Autonomous Agent Systems” on October 31. Its stated starting point is that conventional agents operate through independent sessions, leaving people responsible for coordinating multiple tasks and synchronizing context.

The architecture combines a two-level design, event-driven operation, asynchronous multichannel interaction, separation of storage and computation, and an external watchdog. DeepSeek says these mechanisms transfer session management and context synchronization into the system, allowing people to concentrate on decisions and acceptance of results.

According to the reported technical article, internal Super-agent instances are already running continuously and participating in issue reproduction, fixes, code-review responses, and iteration of DeepSeek Harness itself. When team members report software problems in group chats, the system can identify potential defects and attempt background reproduction and repair without an explicit mention directed at the AI.

A separate QbitAI summary describes proactive submission of pull requests, cross-platform collaboration, parallel complex tasks, and 24-hour operation. The supplied sources provide no benchmark results, task-success rates, inference costs, or public-product availability date.