Phoenix.new – The Remote AI Runtime for Phoenix

Source: fly.io Date: 2025-06-23 URL: https://fly.io/blog/phoenix-new-the-remote-ai-runtime/

Summary

Release announcement for Phoenix.new, a browser-accessible AI coding agent that runs in isolated Fly Machines with root access — no local environment needed. The agent uses a headless Chrome browser to validate UI changes in real-time and inherits full Fly.io infrastructure (networking, port-forwarding, shareable URLs) immediately on app creation. The framing: agents need real computers, not sandboxes.

Implications

Machines API as agent-runtime substrate. Phoenix.new is the clearest example yet of fly.io’s thesis that VM-per-agent workloads are the natural unit of compute for the agentic era. By giving agents root access to isolated Firecracker VMs, they sidestep the sandbox permission friction that constrains tools like Cursor or Codex CLI. The headless Chrome loop — generate, render, observe, iterate — is essentially the same pattern as Playwright-in-the-loop agentic testing. This puts pressure on Vercel’s AI SDK approach (ephemeral serverless) and Render’s PaaS model, both of which lack the “agent has a persistent, rootful machine” affordance. Watch whether other platforms follow with VM-native agent runtimes.