I am impressed by Gemma 4, there’s a lot of power for an on-device model at fast speeds. But I am not convinced you can get real agentic workflows out of a small model on device. So much depends on model judgement, self-correction, and accuracy. Small models are too weak there.