Stop Killing Sandboxes: Agents Need Durable Disposable Computers

Fly.io introduces Sprites: durable, disposable cloud computers that boot in seconds, auto-idle, provide Anycast URLs, and support instant system checkpoints and restores. Kurt Mackey argues agents don’t want stateless sandboxes; they need persistent machines to avoid rebuilds, externalized state, and timeouts, and to leverage full app lifecycles. He demonstrates this with real examples, concluding that durable disposable computers should replace ephemeral sandboxes.

Key Points

• Ephemeral, read-only sandboxes hinder agent workflows by forcing rebuilds, externalized state, and short timeouts; agents need durable computers, not stateless containers.
• Sprites are fast, durable cloud computers: 1–2s creation, 100GB persistent storage, Anycast HTTPS, auto-idle billing, and near-instant system checkpoints/restores.
• Persistence simplifies development: avoid re-installing dependencies and hacking around state with buckets or plan files; treat rollback like git for the whole system.
• Agents can leverage full app lifecycles when state and logs persist, exemplified by Phoenix.new automatically observing and fixing errors from generated apps.
• Real-world illustration: the author built and runs a personal MDM on a Sprite with Claude; for many small apps, dev is prod, and Sprites make continuous, agent-led iteration viable.

Sentiment

The discussion is predominantly positive and curious. Most commenters are genuinely intrigued by the concept and several have already signed up to try Sprites. The Fly team's extensive engagement adds credibility. However, a vocal minority questions the novelty of the offering and raises legitimate reliability concerns based on experience with Fly's other products. The overall tone is constructive rather than hostile — even skeptics frame their objections as wanting to understand the differentiation rather than dismissing the product outright.

In Agreement

• Sprites solve two important problems at once: safe sandbox environments for coding agents in YOLO mode, and a programmable sandbox API for running untrusted code
• The combination of instant creation, automatic scale-to-zero, persistent storage, and fast checkpointing makes Sprites genuinely useful for exploratory computing and personal software
• Full KVM VM isolation is a meaningful security improvement over containers that share a kernel with untrusted cotenants
• The personal software vision — where small apps run on their own Sprite with SQLite and dev equals prod — resonates as a practical paradigm for day-to-day tools
• Claude Code being pre-installed and already knowing how to drive Sprites (creating checkpoints, registering services) demonstrates strong AI-agent integration

Opposed

• Many commenters struggle to see how Sprites differ from Docker containers, LXC, or a traditional VPS with snapshots — the novelty seems limited to better UX on familiar infrastructure
• Fly.io's existing products have reliability issues including API downtime, billing bugs for destroyed instances, and slow support responses, raising trust concerns for a new product
• Several users want local or self-hosted solutions rather than paying for a cloud service, especially when they have powerful local hardware sitting idle
• Vendor lock-in concerns around proprietary checkpoint formats and a single-vendor dependency with no clear migration path
• Practical limitations like low concurrent sprite limits, single-region availability causing latency, and rough documentation undermine the polished pitch