OpenAI Agents SDK Update Adds Secure Sandbox Execution
OpenAI has released a major update to its Agents SDK, adding native sandbox execution for AI agents. As a result, agents can now run in isolated environments with dedicated filesystems, shell access, and cloud storage integration. This update shifts agent workflows from prompt-based interactions to structured execution environments.
Moreover, the new SandboxAgent class creates a clear boundary between orchestration and compute layers. Consequently, agents can perform tasks such as running shell commands, managing files, and installing dependencies within controlled workspaces. In addition, these environments support persistent state across sessions, enabling longer and more complex workflows.
Flexible Storage and Multi-Provider Support
The update also introduces a Manifest system for defining workspace content. Therefore, developers can mount external storage from services like AWS S3, Google Cloud Storage, Azure Blob Storage, and Cloudflare R2. At the same time, the SDK supports multiple sandbox providers, including Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel.
Because of this flexibility, developers can switch providers without modifying agent definitions. Furthermore, built-in capabilities include shell execution, file editing, and context management for long-running tasks. Additional features support memory persistence, skill loading, and port exposure for running services.
Enhanced Security and Persistent Memory
The update strengthens security by separating control logic from execution environments. As a result, sensitive credentials remain protected, while sandbox environments operate with limited permissions. In addition, a snapshot system allows developers to save and restore workspace states during extended tasks.
Meanwhile, the introduction of sandbox memory enables agents to retain knowledge across sessions. This system stores structured data that future runs can reference, improving efficiency over time. Currently, these features are available in the Python SDK, while broader support and additional capabilities remain in development.
