What It Does
Sprites is Fly.io’s persistent VM product built on Firecracker microVMs, designed for AI agent workloads that need state to survive between sessions. Unlike E2B’s ephemeral model (sandbox destroyed after use), a Sprite persists its filesystem to durable object storage, auto-sleeps when idle (no billing), and can be resumed instantly. The signature feature is checkpoint/restore: capture the entire disk and CPU state in ~300ms, then roll back later — described as “git for the whole system.”
Sprites come with Claude pre-installed by default, signaling Fly.io’s explicit focus on the AI coding agent market. Storage is backed by fast NVMe with asynchronous writes to object storage for durability.
Key Features
- Persistent Firecracker microVMs: Hardware-level VM isolation with state that survives between sessions — eliminates package rebuild cycles
- Checkpoint/restore in ~300ms: Capture full disk + CPU state; restore puts everything in place and restarts in under 1 second
- Auto-sleep on idle: Billing stops when idle; state preserved. Eliminates idle charges without losing environment
- Object-storage-backed ext4 filesystem: Up to 100GB persistent storage per Sprite, backed by durable object storage
- Pay-per-use billing: $0.07/CPU-hour, $0.04375/GB-hour memory; approximately $0.44 for a 4-hour session
- Claude pre-installed: Default AI agent integration for coding workflows
- SOC2 certified: Via Fly.io’s existing compliance (Fly.io is SOC2 Type II)
- 1-12 second creation time: Slower than ephemeral platforms but includes full persistent environment setup
Use Cases
- Persistent AI coding agent environments: Agents that install packages, configure tools, and build state over multiple sessions without rebuilding each time
- Checkpoint/rollback experimentation: Agent tries a risky refactoring; checkpoint before, rollback if it fails — git-like workflow for entire system state
- Cost-optimized long-running agents: Agents that work intermittently over days or weeks, with auto-sleep eliminating idle costs
- Mobile development sandboxes: Sprites has documented use for mobile development environments (per Sprites blog)
Adoption Level Analysis
Small teams (<20 engineers): Good fit. Pay-per-use with auto-sleep means you only pay for active compute. The persistent model eliminates time wasted rebuilding environments. Fly.io’s developer experience is well-regarded. No infrastructure to manage.
Medium orgs (20-200 engineers): Good fit. SOC2 compliance via Fly.io. Per-second billing scales predictably. The checkpoint/restore feature is valuable for reproducible agent workflows. However, 1-12 second creation time makes it impractical for high-throughput batch evaluation (use E2B instead).
Enterprise (200+ engineers): Moderate fit. SOC2 certified. However, no BYOC/VPC deployment option — all Sprites run on Fly.io infrastructure. No GPU support. For enterprises requiring VPC deployment, Northflank is a better fit.
Alternatives
| Alternative | Key Difference | Prefer when… |
|---|---|---|
| E2B | Ephemeral-only, Firecracker, 150ms creation | You need high-throughput batch execution and maximum security (no state leakage) |
| Runloop | Persistent with SWE-bench integration and custom hypervisor | You need built-in agent benchmarking and faster vCPUs |
| Daytona | Docker-based, 90ms creation, Computer Use | You need faster creation and browser/desktop automation |
| Northflank | Enterprise VPC deployment, GPU support, ephemeral + persistent | You need VPC/BYOC and GPU workloads |
| CodeSandbox SDK | VM forking with snapshot/hibernation | You are already in the CodeSandbox/Together AI ecosystem |
Evidence & Sources
- Sprites.dev official site
- Fly.io blog: Design and Implementation of Sprites
- Simon Willison: Fly’s new Sprites.dev
- SDxCentral: Fly.io debuts Sprites
- DevClass: Fly.io introduces Sprites
- AI Agent Sandboxes Compared — Ry Walker
Notes & Caveats
- 1-12 second creation time: Significantly slower than E2B (150ms) or Daytona (90ms). Not suitable for high-throughput batch evaluation pipelines that spin up thousands of sandboxes.
- State pollution risk: Persistent environments can accumulate stale dependencies, corrupted configurations, or security vulnerabilities from previous sessions. Requires disciplined checkpoint management.
- No GPU support: Cannot run ML training or inference workloads. Use Modal or Northflank for GPU.
- No BYOC/VPC deployment: All Sprites run on Fly.io infrastructure. Not suitable for teams with strict data sovereignty requirements.
- Fly.io platform dependency: Sprites is a Fly.io product, not an independent open-source project. If Fly.io changes direction, pricing, or shuts down the product, there is no self-hosting option.
- Young product: Launched January 2026. Limited production track record outside of Fly.io’s own case studies. Few independent post-mortems or failure reports available.