OpenAI Updates Agents SDK with Native Sandbox Execution
AI-generated image · US National Wire
The updated SDK introduces a model-native harness and sandbox support, allowing agents to execute code and manage files in controlled environments.
OpenAI has released an updated Agents SDK designed to move agent development from simple prototypes to production-viable systems. According to OpenAI News, the update, announced April 15, introduces a model-native harness and native sandbox execution, enabling agents to inspect files, run commands, and edit code within controlled computer environments.
OpenAI News reports that the updated SDK addresses existing tradeoffs in agent development. While model-agnostic frameworks offer flexibility, they may not fully utilize frontier model capabilities; conversely, managed agent APIs can constrain where agents run and how they access sensitive data. The new SDK aims to align execution with the natural operating patterns of frontier models to improve reliability, particularly for long-horizon tasks coordinated across various tools.
The native sandbox execution allows developers to provide agents with a workspace to install dependencies and run code safely. OpenAI News notes that developers can utilize built-in support for several providers, including Vercel, Modal, E2B, Daytona, Cloudflare, Runloop, and Blaxel, or bring their own sandbox. To ensure portability across these providers, the SDK introduces a Manifest abstraction.
Beyond sandboxing, the updated harness includes several new primitives and tools, as reported by OpenAI News:
* **Filesystem and Code Tools:** Codex-like filesystem tools, a shell tool for code execution, and an "apply patch" tool for file edits. * **Standardized Integrations:** Support for tool use via MCP, progressive disclosure through skills, and custom instructions via AGENTS.md. * **Orchestration:** Configurable memory and sandbox-aware orchestration.
OpenAI News highlighted the practical application of these tools through a sample implementation using the "gpt-5.4" model, where a "Dataroom Analyst" agent was tasked with comparing financial metrics across fiscal years using a local directory.
Early adopters cited the SDK's impact on production workflows. Rachael Burns, Staff Engineer & AI Tech Lead at Oscar Health, told OpenAI News that the updated SDK made it production-viable to automate a critical clinical records workflow. Burns noted that the system's ability to understand boundaries in complex records allowed the company to more quickly understand patient visits.