AI agent survives a reboot: Drives preserves Vercel Sandbox files

R

Guillermo Rauch

@rauchg

Muse, Instinct, OpenClaw, Claude Code… All successful agents have three key parts: 🧠 Brain → the model and the scaffolding, meaning the operating logic 👐 Hands → tools, computer, browser 🗃️ Files → memory, skills, repositories The «simple» route is to assemble all of this in one stateful computer. For example, a Mac mini. That is, you run `claude` or `fx` on your Mac and keep them running all day with `caffeinate`. It has storage, CLI tools, and apps installed. But if you want to run agents in the cloud without unnecessary costs, these parts have to be split apart. 🧠 Agent logic can run in Fluid compute. To reliably survive restarts, deployments, and failures, the event log is made durable through Workflow. 👐 Hands can be moved to a separate fleet of browsers like Browserbase/Kernel, a computer like Sandbox, and even lighter tools like just-bash. 🗃️ 🆕 What was missing was a way to split out storage too. Imagine a nightly cron job that brings an agent's memory together, a «dream». It reads and writes files directly, without starting an entire agent computer. Today we introduce the perfect complement to Sandbox: Drives. We released a computer for agents, and now we are giving them an «external drive» they can connect whenever needed. This is an early version, but the capabilities will expand quickly. Splitting these parts into independent blocks not only noticeably reduces costs, but also makes security and auditing much simpler. I would say that otherwise, it is impossible to run a safe agent at all.

20:59 · Sep 23, 2026 · 22.5K views