Measuring Before Migrating: A Residency Smoke Test for Local and Remote Agent Endpoints

A developer guide argues that agents often break when model calls move across a network, usually because cold-start latency, outbound payload size, and offline behaviour were never measured beforehand. It proposes a small probe that records these three signals for a pair of endpoints and turns them into a decision about where the model should run. Latency is captured as a distribution, with p50 and p95 reported separately and failures counted rather than averaged away. The probe also audits which content leaves the device and checks whether the agent degrades gracefully when the remote endpoint is unreachable. Both endpoints receive identical request bodies loaded from shared files, so the comparison stays fair and the probe can be rerun after runtime updates.