What it does
Hardware-isolated sandboxes for running code safely. Positioned explicitly for AI-generated code and agents — each Sprite gets dedicated CPU, memory, and networking so one agent can't starve another, and execution is contained away from the host.
Hero copy (fly.io homepage): "Run AI-generated code in complete safety." "Agents run better on Sprites — dedicated CPU, memory, networking per instance." "agents that scale to millions of users."
Target ICP
Second ring — AI / agent builders specifically, not Fly's traditional framework-deployment audience. The pitch language ("sandbox," "AI-generated code," "agents") is taken directly from the E2B / Daytona / Modal category.
Inputs / Outputs
- Input: code (from an agent, LLM, or user) + execution environment spec.
- Output: isolated compute environment with CPU/memory/network, returning logs and results.
the API shape — whether Sprites is a distinct API from Fly Machines or a mode flag on Machines. Landing page pitches it as a named product; docs don't yet have a Sprites section at the top level. Could be both marketing wrapper on Machines + its own abstraction.
How it works
Hardware-isolation (not container-isolation). Same substrate as Fly Machines — Firecracker-style microVMs — but the positioning frames the isolation guarantees for the untrusted-code case. Per-instance dedicated resources, in contrast with shared-VM sandboxes.
Two execution modes mentioned: - Persistent — long-lived sandboxes that hold state across calls. - Ephemeral — one-shot invocations, torn down after the run.
Pricing / limits
No Sprites-specific pricing on the page. Presumably uses the same metered Machines pricing. Quotas and limits not documented on the landing page.
UX entry points
- Fly homepage — "Check Em Out!" CTA (primary after hero).
- Footer — "Spin Up a Sandbox" CTA.
No captured screenshots of the Sprites console, dashboard, or flyctl sprites command shape.
Is Sprites invoked via a specific SDK, the Machines API, or a dedicated surface? Is there an HTTP API for agents to call from inside another Fly app?
Differentiators (claimed)
- Hardware isolation — vs. container-based sandboxes (common in the competitor set).
- Dedicated per-instance resources — no noisy-neighbor on CPU or RAM.
- Scale claim — "agents that scale to millions of users." (Not yet benchmarked in captured content.)
Integrations
Not captured. The competitor set (E2B, Daytona, Modal, Runloop) typically ships SDKs for Python + Node and is referenced from agent frameworks (OpenAI Agents SDK, LangChain, LangGraph). Whether Sprites is in those defaults is unknown.
Is Sprites in the OpenAI Agents SDK default provider list? As of the CreateOS internal research (Addendum 3), the default list was Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, Vercel. Fly.io / Sprites was NOT on that list. Worth verifying current state.
Why this matters for CreateOS
Sprites is the most direct competitor to CreateOS's agent-sandbox offering surfaced in this ingest. Not Replit. Not Anything.com. If CreateOS is "the only agent sandbox where the agent can earn," Sprites is a real-time contender for the slot, minus the earning primitive. The differentiator CreateOS has — MPP + on-chain payments — is exactly what Sprites cannot replicate without entering money-movement.
Sources
- fly.io homepage — Sprites section