OpenAI adds sandbox execution to Agents SDK

OpenAI updated its Agents SDK to include native sandbox execution and a model-native harness that run automated workflows inside isolated environments. The changes add snapshotting and rehydration and separate the control plane from compute to reduce exposure of credentials and sensitive data.

The SDK introduces a Manifest abstraction that defines a predictable workspace. Teams can mount local files, set output directories and connect to enterprise storage providers such as AWS S3, Azure Blob Storage, Google Cloud Storage and Cloudflare R2. The Manifest limits where agents read and write data, allowing governance teams to trace where inputs came from and how outputs were produced.

New runtime features include configurable memory, sandbox-aware orchestration, filesystem tools inspired by earlier coding models, and primitives for integrating external tools, custom instructions and file edits. Developers can run sequences of skills and code through a shell tool to perform multi-step tasks and to progress tasks in stages.

OpenAI provides out-of-the-box sandbox support so engineering teams do not need to build their own execution layers. Organizations can deploy custom sandboxes or use built-in integrations with providers including Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop and Vercel. The update separates the control harness from the compute layer so credentials remain outside the runtime environments where model-generated code executes.

Snapshotting and rehydration save and restore environment state so runs can resume from the last checkpoint after a container crash or timeout. The SDK can route subagents into isolated environments, run many sandboxes in parallel to speed execution, and scale resources dynamically to match workload.

Oscar Health tested the updated SDK to automate a clinical records workflow that had failed under previous approaches. According to Rachael Burns, staff engineer and AI tech lead at Oscar Health, the SDK made the workflow production-viable by allowing the system to extract accurate metadata and identify encounter boundaries in long medical records, which shortened the time to parse patient histories and supported care coordination.

The new capabilities are generally available through OpenAI’s API with standard pricing based on tokens and tool use. The initial rollout targets Python developers, with TypeScript support planned for a later release. OpenAI also plans additional features for both language libraries and broader support for more sandbox providers over time.