DeepSeek might have just revealed its approach to AGI

According to leaked internal discussions, DeepSeek sees continual learning as one of the biggest remaining hurdles to AGI.

Then it released DeepSeek Harness, and quietly rewrote the base agent runtime layer.

Its new runtime makes almost everything hot-swappable in a long-running session: tools, skills, memory, execution logic, even the model itself.

We usually assume the unit of continual learning is the model. Maybe DeepSeek is thinking one level higher. Maybe the thing that learns continuously is not the model, but the agent, while the model is just one component the agent can replace, specialize, or eventually retrain as it evolves.

If so, this is not merely a new agent harness. It may be a glimpse of DeepSeek’s path toward AGI.