Introducing Temporal and agentic sandboxes: the OpenAI Agents SDK

OpenAI recently introduced sandbox support within its Agents SDK, enabling agents to execute code, manipulate files, and run shell commands in isolated environments using providers like Modal, Daytona, Docker, and E2B. This new feature, developed with input from OpenAI's engineering team, includes a Temporal extension that allows sandbox agents to function as durable Temporal workflows, maintaining state and reducing resource waste by idling without consuming compute resources. The demo provided showcases the architecture consisting of AgentWorkflow, SessionManagerWorkflow, and a Textual-based terminal interface (TUI), highlighting capabilities such as session forking and backend switching while preserving workspace state. This advancement ensures agents are more reliable and production-ready through enhanced fault tolerance and state management, paving the way for scaling to thousands of concurrent sessions without idle sandbox costs. The integration with Temporal provides durability and persists the workflow state across restarts, with plans for further integrations with platforms like Pydantic AI and Amazon Bedrock Strands.