ChinaDeepSeek’s agent design finished half its tasks in 12.4 minutes.
It runs many tasks at once because human attention, not the model, is now the bottleneck.
A Router session only dispatches, isolated Worker sessions do the work, and an external Watchdog session checks their progress.
The tests used synthetic workloads and more tokens, so they are a preliminary reference, not a product verdict.
DeepSeek has published the prototype on GitHub but has released no product.
How each outlet framed it
- Zhidx
- reads DeepSeek's Super-agent as trading tokens for time and stability, flagging the experiments as synthetic-load prototypes rather than product results
Sources: Zhidx