July 2026 Summaries

Sailboxes, now generally available after a beta period, are cloud-based, kernel-isolated Linux virtual machines designed for long-horizon AI agents that need persistent state, Docker support, root access, local NVMe storage, and unrestricted runtimes. Sail states that the service costs more than 70% less than competing sandbox providers by billing only active CPU, memory, and disk usage rather than reserved capacity, while allowing machines to sleep during Sail inference calls. Its efficiency relies on transparent live VM migrations that optimize hardware utilization, accepting occasional seconds-long latency increases in exchange for lower costs. Sailboxes are being used for persistent cloud agents, parallel scientific experimentation through Quadrillion Labs’ Qualia Cloud, reinforcement-learning rollouts using asynchronous inference, and development workflows such as coding agents, bug detection, and automated triage. Users can create and connect to Sailboxes through the Sail CLI, with further setup information available in the company’s documentation and Slack community.

Sailboxes is a cloud platform for long-running agents that uses live migration to reduce costs by charging only for active resource use and operating workloads on ephemeral, load-balanced infrastructure. Its migration system preserves open network connections through a persistent Layer 4 proxy and in-VM sidecar, which also buffers packets during brief outages and supports functions such as egress controls, credential injection, and SSH continuity. To move large virtual machines quickly, Sailboxes uses Linux user-space page fault handling to start a VM on a new host with minimal state, retrieve needed memory pages from the previous host or S3, and improve transfer performance with 2 MB memory pages. Firecracker microVMs, virtio-mem hot-plugging, memory-pressure monitoring, and background checkpoints allow instances to expand or shrink memory dynamically while reducing migration data. The company says these mechanisms enable migrations with only seconds of disruption, support pricing of $0.015 per vCPU-hour, and kept a Minecraft server operating through repeated forced migrations, though migrations can add brief latency and are intended to occur only occasionally for host reclamation or fleet balancing.