The Design & Implementation of Sprites

Source: fly.io Date: 2026-01-15 URL: https://fly.io/blog/design-and-implementation/

Summary

Engineering writeup and release announcement for Sprites, Fly.io’s “disposable cloud computers” — Linux VMs designed for instant creation, persistent storage, automatic sleep, and near-zero idle cost. The post explains three deliberate inversions of Fly Machines’ design: no container images (standard base only), object storage for disks (S3-backed rather than NVMe), and inside-out orchestration (most orchestration runs inside the VM, not on the host).

Implications

Machines API as agent-runtime substrate. Sprites are explicitly optimized for interactive, short-lived agent workloads — the same profile as MCP tool invocations and agentic coding loops. By removing OCI image overhead and backing storage in S3, Fly positions Sprites as cheaper and faster cold-starts than standard Machines, which directly pressures the “run a container per agent task” approach used by competitors. This connects to the broader “Code and Let Live” / “Our Best Customers Are Now Robots” thread — the platform is actively reshaping around agent consumption patterns. Watch whether Sprites become the substrate Sprites-speak-MCP sits on.