Vercel adds Agents Sandbox
Vercel integrated Sandbox with the OpenAI Agents SDK to run isolated microVMs that handle code, files and data for agent workflows, according to its announcement. (x.com)
An AI agent that writes code needs a locked room to run it in, and Vercel says its Sandbox now plugs into OpenAI’s Agents SDK for that job. (openai.com) OpenAI said on April 15 that the Agents SDK now supports native sandbox execution, letting agents work across files and tools while running code in an isolated environment. Vercel’s Sandbox docs describe those environments as ephemeral Linux microVMs created on demand. (openai.com) (vercel.com) A microVM is a stripped-down virtual machine, like a disposable computer that starts fast and disappears after the task ends. Vercel says Sandbox can run commands, manage files, take snapshots, and block access to secrets, databases, cloud resources, and SSH keys. (vercel.com 1) (vercel.com 2) The setup targets a basic problem in agent design: models can plan steps, but many real tasks require executing code, editing files, or using software without touching production systems. OpenAI’s Agents docs say the SDK is for applications that own orchestration, tool execution, approvals, and state across multi-step work. (openai.com 1) (openai.com 2) OpenAI already offers its own Code Interpreter tool for Python work in a sandbox and a computer-use tool for operating software through screenshots and interface actions. The new SDK update moves that pattern into a broader agent framework that outside infrastructure providers can hook into. (openai.com 1) (openai.com 2) (openai.com 3) For Vercel, the announcement extends a product it began rolling out in 2025 and pushed to general availability in February 2026. The company says Sandbox is built for untrusted code, including AI-generated output, user uploads, and third-party code. (vercel.com) (vercel.com) Vercel says Sandbox runs on Firecracker-based microVMs and on its Fluid compute platform, with features such as filesystem snapshots, parallel instances, runtime network controls, and container support. Those details matter for longer agent jobs that need to resume work, fan out tasks, or limit what external services they can reach. (vercel.com) (vercel.com) (vercel.com) OpenAI’s own examples now point developers toward coding agents that can scaffold apps, edit code, and execute commands across full codebases. Vercel’s pitch is that those steps can happen inside a disposable machine instead of on the app’s real servers. (openai.com) (vercel.com) The result is a cleaner split of labor: OpenAI supplies the agent framework and models, while Vercel supplies the execution box where risky work can run and then vanish. That is the part companies have been asking for as agents move from chat windows into production workflows. (openai.com) (vercel.com)
