OpenAI Agents SDK Major Update Adds File and Code Execution in Sandboxes

ME News report, April 16 (UTC+8): According to monitoring by Beating, OpenAI has released a major update to its Agents SDK. The core change provides agents with a genuine working environment: agents can now read and write files, install dependencies, execute code, and invoke tools within a controlled sandbox, rather than being limited to conversational interactions. The SDK’s agent execution framework (harness) now includes configurable memory, sandbox-aware orchestration, and filesystem tools derived from Codex (shell command execution and apply patch code editing). The framework also integrates multiple emerging industry-standard agent protocols: MCP for tool invocation, AGENTS.md for custom instructions, and Skills for progressive capability declaration. Developers no longer need to assemble these infrastructure components manually. Sandbox execution is another key focus of this update. The SDK natively supports seven sandbox providers: Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel, with the option for developers to integrate their own sandbox environments. The newly introduced Manifest abstraction uniformly describes an agent’s workspace, enabling mounting of local files, defining output directories, and connecting to storage services such as AWS S3, Google Cloud Storage, Azure Blob Storage, and Cloudflare R2—all using the same configuration from local prototyping to production deployment. Architecturally, the SDK decouples agent scheduling logic from the compute environment. This offers three key benefits: credentials never enter the execution environment of model-generated code, reducing risks of prompt injection and data leakage; agent state is externalized, enabling snapshots and recovery so that sandbox container crashes do not result in lost progress; and multiple sub-agents can execute tasks in parallel across separate containers. The new features are now available to all API users and are billed according to standard token usage and tool invocation rates. Currently, only Python is supported; a TypeScript version will be released in a future update. (Source: BlockBeats)