Long-horizon agent tasks often depend on frontier models running in the cloud to achieve strong results. The catch, as the EdgeWise preprint notes, is that invoking these models for every step of a trajectory adds up quickly in inference cost. EdgeWise is designed around a simple premise: not every step needs the full power of a frontier model.

Instead of sending the entire trajectory to the cloud, EdgeWise selectively delegates only certain steps, keeping the rest of the work at the edge. This could make long-horizon agents more economical to run while preserving the performance that comes from cloud-based reasoning. The abstract does not yet detail how the system decides which steps to delegate.

Because this is a single arXiv abstract, there are no independent results or contrasting sources to compare. The contribution, as stated, is the selective-delegation approach itself.