TL;DR
Fly.io is pitching a different approach to running code for AI agents: durable, fast-to-create virtual machines called Sprites that keep state, support instant checkpoints and restores, and stop billing when idle. The company argues ephemeral, read-only sandboxes are a poor fit for agents that benefit from persistent storage and longer-lived compute.
What happened
Fly.io unveiled a runtime they call Sprites: small Linux hosts that appear in about one to two seconds, provide a root shell, and start with a 100GB filesystem. Users can install packages, create instant checkpoints, and restore entire system state in about a second. Sprites automatically go idle and stop metering when unused, which makes it economical to run many of them. The platform exposes Sprites via Fly’s Anycast network (so each can get an HTTPS URL) and emphasizes durability — a Sprite keeps its filesystem and runtime state until the owner deletes it. Fly.io positions Sprites as an alternative to traditional ephemeral sandboxes and stateless containers, arguing that agents (like Claude, cited in the post) and other workflows benefit from persistent machines with fast snapshot/restore capabilities. The blog also describes real-world use: a personal MDM built and run continuously on a Sprite and integrations used by developer tooling.
Why it matters
- Agents and interactive developer workflows often need durable state; Sprites avoid rebuilding environments on each run.
- Instant checkpoints and fast restores let users recover from mistakes quickly and use snapshots as part of normal workflows.
- Automatic idling and stop-metering make maintaining many long-lived machines cost-effective compared with always-on instances.
- Having durable local storage reduces the need to architect external infrastructure for state (S3, Redis, etc.) just to work around sandbox ephemerality.
- Sprites enable new agent-driven patterns where development, testing, and production can coexist on the same machine.
Key facts
- Sprites boot into a usable root shell in roughly 1–2 seconds.
- Each Sprite begins with about 100GB of capacity by default.
- Users can create checkpoints instantly; Fly.io reports checkpoint creation completes immediately and restores take about one second.
- Sprites go idle and stop metering automatically when inactive.
- Sprites are exposed over Fly’s Anycast network and can receive HTTPS URLs.
- Fly.io contrasts Sprites with its existing Fly Machines (stateless containers) and with typical ephemeral, read-only sandboxes.
- The company argues industry efforts to snapshot ephemeral sandboxes cost tens of millions in engineering resources.
- Fly.io warns Sprites are not intended as the vehicle for applications serving millions of users.
What to watch next
- Whether other cloud providers or runtimes adopt persistent, checkpointable short-lived VMs similar to Sprites — not confirmed in the source.
- How tooling and orchestration evolve to manage large fleets of durable, low-cost Sprites at scale — not confirmed in the source.
- If agents will be granted broader hardware access (sound cards, USB devices) in these environments — not confirmed in the source.
Quick glossary
- Sprite: A Fly.io offering described as a durable, quickly created Linux host with built-in checkpoint and restore capabilities.
- Agent: An automated system or model (for example, Claude) that performs tasks, interacts with environments, and may benefit from persistent state.
- Sandbox: An isolated execution environment often used to run untrusted code; commonly ephemeral and read-only in many agent deployments.
- Checkpoint / Restore: Saving the entire system state at a point in time and later restoring a machine back to that saved state, enabling fast recovery or rollback.
- Anycast: A network addressing method where a single IP address is advertised from multiple locations, often used to provide low-latency access and HTTPS endpoints.
Reader FAQ
What is a Sprite?
A Sprite is a durable, fast-to-start Linux host from Fly.io that supports instant checkpoints and restores, provides persistent storage, and can stop metering when idle.
How are Sprites different from typical ephemeral sandboxes or containers?
Sprites keep filesystem and runtime state across sessions, can be snapshotted and restored quickly, appear in 1–2 seconds, and idle to stop billing; containers are typically stateless and ephemeral.
Are Sprites intended for apps serving millions of users?
The source says you wouldn’t want to ship an app to millions of people on a Sprite.
Do Sprites stop charging when idle?
According to the source, Sprites go idle and stop metering automatically.
Will Sprites give agents access to hardware like sound cards or USB?
Not confirmed in the source.
Image by Annie Ruygt The state of the art in agent isolation is a read-only sandbox. At Fly.io, we’ve been selling that story for years, and we’re calling it: ephemeral…
Sources
- Code and Let Live
- Fly's new Sprites.dev addresses both developer …
- Sprites – Stateful sandboxes
- Top Sandbox Platforms for AI Code Execution in 2025
Related posts
- Finding and Fixing Ghostty’s Largest Memory Leak — Root Cause and Patch
- My CES 2026 roundup: Apple-adjacent accessories for iPhone and Mac users
- These standout smart-home gadgets from CES 2026 that impressed me