OpenAI adds native sandbox execution to Agents SDK

OpenAI has added native sandbox execution and a model-native harness to its Agents SDK. The features are available through the API, are priced by tokens and tool use, and begin with Python support; TypeScript support is planned for a later release.

The update standardizes the infrastructure for running autonomous programs by aligning execution with how large models operate. The SDK now includes controls for isolated workflows, credential separation, filesystem-style tools and tools for managing state so agents can coordinate across systems and work with unstructured data.

Before this release, teams often chose between model-agnostic frameworks, provider SDKs and managed agent APIs. Model-agnostic frameworks offered flexibility but did not always use the latest model capabilities. Provider SDKs exposed model internals without strong governance controls. Managed agent APIs limited where systems could run and how they accessed sensitive data. OpenAI’s changes introduce a single, model-native harness intended to standardize those deployment patterns.

The harness provides configurable memory, orchestration aware of sandboxes, and tools similar to a filesystem to reduce custom connector work. Standard primitives listed in the SDK include tool calls via MCP, custom instruction files through AGENTS.md, file edits through an apply-patch tool, and sequential task execution through skills and shell-based code execution.

The SDK adds a Manifest abstraction to define a workspace with mount points and output directories. Teams can connect those workspaces directly to enterprise storage such as AWS S3, Azure Blob Storage, Google Cloud Storage and Cloudflare R2. The Manifest gives the model clear boundaries for locating inputs and writing outputs, limiting access to specified data locations and helping governance teams track provenance from prototype to production.

Native sandbox execution provides an environment where code runs with only the files and dependencies it needs. Customers can run their own sandboxes or use built-in support for providers including Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop and Vercel. The SDK separates the control harness from the compute layer so credentials and central control-plane access remain outside the execution environments where model-generated code runs.

The SDK externalizes system state and supports snapshotting and rehydration. If a sandbox container crashes or an API times out, teams can restore state into a fresh container and resume from the last checkpoint rather than restarting long processes from the beginning. The separated architecture also supports dynamic resource allocation: runs can invoke one or multiple sandboxes, route subagents into isolated environments and parallelize work across containers.

OpenAI describes the release as generally available to customers through the regular API without custom procurement. The company plans future SDK features such as code mode and subagents and intends to expand support for additional sandbox providers and integration methods.

Oscar Health tested the updated Agents SDK to automate a clinical records workflow that previous approaches could not handle reliably. The insurer used the SDK to extract accurate metadata and identify the boundaries of individual patient encounters inside long medical records. Rachael Burns, staff engineer and AI tech lead at Oscar Health, described the SDK as “production-viable for us to automate a critical clinical records workflow that previous approaches couldn’t handle reliably enough.” She added the system helps staff more quickly understand what happened during each patient visit, speeding care coordination and improving member experience.