DeepSeek Harness Team Details a Continuously Running Super-agent

The DeepSeek Harness team has a continuously running internal instance built on a Super-agent architecture designed for multi-task and team collaboration, helping reproduce and fix issues and respond to code reviews.

The architecture uses two layers, event-driven operations, asynchronous communication across multiple channels, separate storage and compute, and an external Watchdog. The system manages sessions and synchronizes context, allowing people to focus on higher-level decisions and validation. The internal instance also contributes to ongoing development of DeepSeek Harness.

The Super-agent label describes the agent form and its architecture; the running instance is built on that approach. Mainstream agent designs struggle to support work across multiple tasks and teams at scale because sessions are often treated as independent. In that model, the scope of a single task depends on model capability, while a person’s attention limits how many tasks they can manage.