The paper argues that the state-of-the-art read-only sandbox for agent isolation is obsolete and that "ephemeral sandboxes are obsolete." Instead, it introduces "Sprites," which are presented as "disposable cloud computers" designed to address the limitations of stateless containers for AI agents.

The core methodology revolves around providing agents with durable, fast-booting, and stateful computing environments. A Sprite is described as a root shell on a Linux computer that comes online in 1-2 seconds, comparable to the time it takes to SSH into an existing host. Unlike traditional virtual machines or ephemeral containers, Sprites offer:

1. Fast Creation: A Sprite can be created in approximately 1.0 second using sprite create ..
2. Durable Storage: Each Sprite comes with 100GB of persistent storage, ensuring that installed software (e.g., ffmpeg) and created files remain across sessions, even after periods of inactivity.
3. First-Class Checkpoint and Restore: Sprites support instantaneous state capture and restoration. Users can create a checkpoint of the entire system state using sprite-env checkpoints create and restore to a previous state using sprite checkpoint restore v1. This restoration process takes about one second, making it practical for interactive use, akin to version control for an entire system.
4. Automatic Idling and Metering: Sprites go to sleep when inactive and stop metering, making them cheap to maintain in large quantities.
5. Network Connectivity: Sprites are integrated with the Fly.io Anycast network, allowing them to expose HTTPS URLs.
6. Simpler Management: They operate without Dockerfiles, implying a different underlying orchestration model than typical container deployments.

The paper posits that AI agents ("Claude") do not want stateless containers or sandboxes but rather "computers" that possess durable storage and do not vanish after a single job. This addresses several pain points in current agent development:

• Elimination of Environment Rebuilds: Agents no longer need to rebuild development environments (e.g., node_modules) for every task or PR, saving significant time and resources.
• Simplified State Management: The need for complex external infrastructure (S3 buckets, Redis servers, RDS instances) or "plan files" to persist state is removed, as agents can simply write files and trust them to remain.
• Support for Long-Running Tasks: Current ephemeral sandboxes are typically limited by short time budgets (e.g., 15 minutes). Sprites enable agents to handle compute-intensive or long-duration tasks that involve significant interaction time, like building documentation sites or continuous monitoring of applications.
• Full Application Lifecycle Interaction: Agents can interact with and understand the entire lifecycle of an application, including monitoring logs for exceptions and autonomously working to resolve issues, as demonstrated by the Phoenix.new example.
• "Dev is Prod, Prod is Dev" Paradigm: Sprites enable a model where a piece of software (e.g., a personal MDM application) can be continuously developed and run on the same instance, with agents evolving it as needs change, without the traditional separation of development and production environments.

Technically, Sprites are related to Fly Machines but feature an "entirely new storage stack" and "orchestrated differently," allowing for their unique combination of rapid provisioning, durability, and cost efficiency. The argument concludes that the future for coding agents lies not in read-only sandboxes in Kubernetes clusters, but in fully-fledged, disposable yet durable computing instances.