OpenAI Updates Agents SDK with Sandbox Support for Enterprise AI Development

Apr 15, 2026

OpenAI has updated its agents software development toolkit with native sandbox support and new harness capabilities, enabling developers to build AI agents that operate in controlled, isolated environments. The update, launching first in Python with TypeScript support planned for later, allows agents to access files and execute code within specific workspaces while protecting overall system integrity. Karan Sharma from OpenAI's product team explained that the in-distribution harness provides components for frontier models—the most advanced general-purpose AI systems—to work with approved tools during both testing and deployment. The new capabilities address risks associated with unsupervised AI agents, which can behave unpredictably, by confining their operations to siloed workspaces. Available to all API customers at standard pricing, the SDK expansion represents OpenAI's effort to make enterprise AI agent development safer and more practical for business applications.

OpenAI's Agents SDK Update with Sandbox and Harness Features

▪OpenAI updated its agents software development toolkit (SDK) with new features designed to help businesses create their own agents
▪OpenAI is working to bring agent capabilities like code mode and subagents to both Python and TypeScript
▪OpenAI's agents SDK sandbox integration allows agents to work in a siloed capacity within a particular workspace, accessing files and code only for particular operations
▪OpenAI's updated agents SDK includes a sandboxing ability that allows agents to operate in controlled computer environments
▪OpenAI's updated agents SDK provides developers with an in-distribution harness for frontier models that allows agents to work with files and approved tools within a workspace

Safety and Control Mechanisms for Enterprise AI Agents

▪OpenAI's agents SDK sandbox integration protects the system's overall integrity while allowing agents to access specific files and code
▪Running AI agents in a totally unsupervised fashion can be risky due to their occasionally unpredictable nature
▪An in-distribution harness allows companies to both deploy and test agents running on frontier models
▪Frontier models are considered to be the most advanced, general-purpose AI models available
▪In agent development, the harness refers to the components of an agent besides the model that the agent is running on

Rollout Timeline and Platform Support

▪OpenAI plans TypeScript support for the agents SDK harness and sandbox capabilities for a later release
▪OpenAI's new Agents SDK capabilities will use standard pricing
▪OpenAI's new agents SDK harness and sandbox capabilities are launching first in Python
▪OpenAI's new Agents SDK capabilities are being offered to all customers via the API
▪OpenAI said it will continue to expand the Agents SDK over time