Cloud writes the program, a small on-device model answers the questions: AppWorld success rises up to 14.9 points with user data kept local

An agent executing user tasks can split writing the program from reading the user's data: the cloud model only writes the program, and a small local model only answers the bounded questions that come up while that program runs. The two previous paths each carried a cost: agents driven by large cloud-hosted models are the most accurate but read the user's private data, while models small enough to run on edge devices keep that data local but yield far worse results. On AppWorld, this split improved the success rate of a code-only agent in all six combinations of backbone model and test partition, by up to 14.9 percentage points, recovering a substantial part of the gap to a cloud-hosted ReAct loop.

The program is written before any data is seen, so it fails whenever a step depends on interpreting content that appears only at run time. The intermediate architecture has the cloud-written program defer bounded questions over user data to a small local model through an ordinary function call. The evaluation runs on AppWorld across six combinations of backbone model and test partition, with user data kept on the device.

This report draws on summary-level material, and the demonstrated scope is an AppWorld evaluation across six combinations of backbone model and test partition, comparing a code-only agent with a cloud-hosted ReAct loop. It is aimed at researchers and engineers working on agent architectures and on-device deployment.

The gain is stated as up to 14.9 percentage points, and the specific values for the other combinations are not given in the material; the gap to a cloud-hosted ReAct loop is described as substantially recovered rather than closed.