Hosting company Fly.io introduces Sprites, lightweight virtual machines based on Firecracker. These are specially designed to isolate coding agents in their own environment.
CEO Kurt Mackey explains that traditional “ephemeral sandboxes” (read: containers) are not suitable for agents like Claude. Persistent VMs allow agents to resume work on pull requests without having to rebuild the development environment each time. According to Mackey, a Sprite comes online within one to twelve seconds. When inactive, they shut down but retain permanent storage.
Claude is not a “pro developer,” Mackey says. Instead, it is a “hyperproductive five-year-old savant.” In other words, it tries everything, including what works, but “wants to stick its finger in every available socket.” To protect against this (and the end user’s development environment), an agent is given a “computer,” or a VM.
Billing is based on CPU time, memory usage, and storage consumed. Sprites also offer checkpoint and restore functionality, allowing users to quickly roll back a damaged environment.
New design for security
Sprites differ from Fly.io’s standard VMs in that they have a completely redesigned storage architecture and different orchestration. Unlike standard Fly VMs, Sprites are not based on Docker images. Mackey acknowledged on Hacker News that the standard flyctl interface is complex. The Sprite CLI should remedy these issues.
Although Sprites can handle different workloads, Fly.io focuses specifically on agentic coding. The company installs Claude as a standard option. AI coding introduces vulnerabilities such as prompt injection and hallucination, which makes execution on a developer’s local machine risky. Research shows that by June 2025, AI-generated code introduced more than 10,000 new security issues per month at large organizations. DevClass highlights a painful example of the danger posed by AI-driven coding assistance. A Google Antigravity user lost their entire D drive because the AI tool “accidentally” deleted it.
Growing security concerns
Even within isolated environments, agents remain vulnerable, often due to API access to systems such as GitHub. Fly.io developer Thomas Ptacek announced that the company will soon release an open-source version for local execution. This should offer developers who prefer local environments over cloud providers a safer option than Docker containers or current local tools.
Read also: Exabeam focuses on security of autonomous AI agents