Particle.news
Download on the App Store

OpenAI Rolls Out Agents SDK With Model-Native Harness and Sandboxed Workspaces

A security-first design targets enterprise needs by isolating execution to keep credentials out of model-run containers.

Overview

  • The SDK update, which OpenAI announced Wednesday, is available to all API customers and uses standard token and tool pricing.
  • The architecture separates the agent harness from the compute so tool calls run unprivileged and company secrets remain outside the sandbox.
  • A new model-native harness lets agents handle files, run approved tools, and manage memory inside a defined workspace suited to long tasks.
  • Teams can bring their own containers or use Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, or Vercel, and mount storage from AWS S3, Google Cloud Storage, Azure Blob, or Cloudflare R2.
  • The release starts with Python, TypeScript support is planned, and OpenAI is adding durability features like snapshot and rehydration plus options such as code mode and subagents.