Paper proposes using small local models as interpreters for cloud-written agent programs
The arXiv preprint (arXiv:2610.14209v1) studies a hybrid agent architecture where a cloud-written program defers bounded, runtime-dependent questions over private user data to a small local model via function calls. Evaluated on AppWorld across six backbone/test splits, this approach improved code-only agent success rates in all cases—by up to 14.9 percentage points—closing much of the gap to a fully cloud-hosted ReAct loop while keeping data on-device.