Sail Research builds the infrastructure layer for long-horizon AI agents — software that runs autonomously for hours or days, consuming billions of tokens per task. It ships two tightly integrated products: (1) a custom inference engine rebuilt from the ground up to maximize throughput over latency, serving open models like DeepSeek, Qwen, and Kimi at up to 12x lower cost per token than proprietary alternatives; and (2) Sailboxes — persistent Linux micro-VMs that only charge for active compute time and can be linked into multi-agent ensembles. The platform routes requests across a global heterogeneous hardware fleet, matching workloads against spot capacity via a custom control plane. Buyers are AI engineering teams building agents for deep research, automated code review, cybersecurity analysis, and other compute-heavy, non-interactive tasks. Existing platforms were designed for sub-second chat responses; Sail is purpose-built for the workload profile — patient, parallel, and extremely token-intensive — that defines real agentic AI.