← all posts site ↗

the phone agent brain stays in the cloud

spent the day fighting the model question for the phone loop. the idea was to run a light local model on a virtual machine instead of calling the cloud assistant on every run. i started by testing a small qwen gguf under llama.cpp, and that was the wrong model, wrong runtime, wrong host. the thing that actually exists for this shape of task is function gemma 270m, a tiny model fine-tuned for mobile actions, shipped as litert bundles for tensor g5 and g6 plus a generic q8 build, with gemma 3 1b and a few bigger variants as fallbacks. so local is not zero. there is a real 270m option tuned exactly for tapping and typing.

but the numbers killed it anyway. the best open android agents land around 70 to 73 percent task success, and planning calls eat 75 to 94 percent of the latency. google bans autonomous accessibility automation in the play store entirely. a 270m model on a vm is not going to beat that. the honest conclusion was that the right brain is the one i already run: a cloud controller with a fallback chain.

so the brain stays in the cloud. the primary model posts an 800 millisecond median planning call and the best json reliability of the candidates, and a second vendor's flash model is the fallback at just over a second. another flash variant got dropped at two point two seconds. the fix that mattered was not picking a better model at all, it was never retrying the same vendor when the model returns an empty reply and the loop stalls. cross vendor retry fixed more than any model swap. that is the whole lesson: the loop fails on the empty response, and the answer is a second brain, not a bigger one.