The DeepSeek Harness team published a long technical article on its official WeChat account, "Super-agent Design Exploration: Toward a More Autonomous Agent System," and open-sourced the prototype dsh-superagent-prototype (MIT license). The team explicitly states that this is an engineering-oriented research prototype, not a ready-made product. But one scene in the article is striking—instances already running internally take part in defect reproduction and code review, and even when no one @s it in a group chat, it can sniff out potential bugs and reproduce them in the background.

What stalls multi-tasking is people, not models

Mainstream Agents still operate on a single Session basis, with sessions isolated from one another. When people want to do several things in parallel, they can only repeatedly switch windows and keep context in their heads. The team observed that a person can usually keep up with only 3 to 5 active sessions, creating a vicious cycle: the ceiling for a single task depends on the model, but the bottleneck for multiple tasks gets stuck on human attention. Super-agent is meant to break exactly this barrier.

One scheduling center, a group of working "hands"

The core is a two-layer division of labor: a persistent front-end Router handles interaction and dispatches tasks, and a lightweight model is enough for it; each task gets a context-isolated Worker, runs in parallel with the strongest model, and reports back to the Router when done. Combined with event-driven design, compute-storage separation, and Watchdog supervision, stuck tasks are automatically detected and repaired. In a stress test of about 300 messages containing 20 requests, P50 was 12.4 minutes and P95 was 15.7 minutes, while the single-Session approach reached a P95 of 29.5 minutes.

Cost is even more aggressive—V4-Pro costs about $1.32/$3.96 per million tokens, an order of magnitude lower than GPT-5.5 and Opus 5.5. When the concurrency limit shifts from "human attention" to "system throughput," only then can Agents truly step out of the chat window.