Execution-State Capsules: graph-bound checkpoint/restore for on-device agent serving

arXiv·medium signal

This paper proposes Execution-State Capsules, a graph-bound mechanism to checkpoint and restore full execution state for low-latency, small-batch, on-device 'physical-AI' serving, going beyond the prefix reuse of paged/radix KV caches. The approach targets workloads where high-throughput KV caching is a poor fit, such as edge and embodied agents that need to suspend and resume mid-task. For builders deploying agents on constrained devices, it points toward faster cold-starts and resumable agent sessions.