OpenAI has pushed a significant update to its Agents SDK, adding native sandbox support so AI agents can operate in isolated environments — separate from the control logic that orchestrates them. The idea is straightforward: if an agent breaks something or hits a failure, it spins up in a fresh container and picks up where it left off. Less blast radius, easier recovery.
What's new
The updated SDK bundles Model Context Protocol (MCP) for tool usage, a shell tool for running commands, an apply-patch tool for file edits, and AGENTS.md support for custom instructions. A manifest function handles workspace definitions and works with local files as well as cloud storage — AWS S3, Google Cloud Storage, and Azure Blob Storage are all supported. For sandbox providers, OpenAI is working with Cloudflare, Vercel, E2B, and Modal, though developers can wire in their own sandbox implementations. Python support is live now; TypeScript is coming. Standard API pricing applies.
Why it matters
Sandboxing has been a persistent gap in agent frameworks — agents that can execute shell commands and edit files are genuinely risky without hard environment boundaries. By separating execution environment from control logic, OpenAI is addressing one of the more practical concerns developers have raised about deploying long-running agents in production. It also makes scaling cleaner: stateless containers are easier to manage than persistent, potentially contaminated environments.
What to watch
The TypeScript rollout is the near-term milestone to track, since most front-end and full-stack developers building on the SDK will be waiting on it. Longer term, how well the manifest and sandbox abstractions hold up across different cloud providers — and whether the MCP integration becomes a standard pattern for tool-calling in production agents — will determine if this update has real staying power or just looks good on a changelog.