Why Free Sandbox Replicas Cannot Be Trusted as Durable Storage for AI Planners

A software architecture analysis warns that AI planner systems risk data loss when they treat free sandbox environments as reliable storage, since these replicas can be preempted at any time without warning. The core problem arises when a planner commits a saga step after receiving an HTTP 200 response, not realizing the replica that accepted the write may no longer exist on the next operation. The author proposes using a fencing token tied to a sandbox epoch — a monotonic integer representing the current sandbox generation — so that a write is only committed when both a receipt and a live epoch are confirmed together. Without this mechanism, a retry after preemption can silently produce a duplicate write, turning what appears to be a model reliability issue into an undetected protocol failure. The article frames this as an architectural constraint relevant specifically to systems using free or preemptable compute, rather than reserved virtual machines with controlled storage.