From model to agent: Equipping the Responses API with a computer environment

OpenAI detailed in a blog post how developers can build an agent runtime using the Responses API, a shell tool, and hosted containers. The system allows agents to execute shell commands, manage files, and maintain state across sessions, all within a secure, scalable environment. According to OpenAI, this moves beyond simple model inference to an agent-centric architecture where the runtime handles tool execution and context. The setup separates agent logic from the underlying model, enabling multiturn interactions and complex workflows like code editing, data processing, or system administration. For builders, this means they can now deploy autonomous agents that interact with a computer environment without managing low-level infrastructure. The approach leverages OpenAI's existing Responses API for structured communication and adds a hosted container for sandboxed execution. Practical implications include automated testing, server management, and custom tool chains—all orchestrated through the API. The post emphasizes reproducibility and safety, with containers that reset after each session. This development gives developers a reference architecture for creating agents that go beyond chat to perform real actions on a remote machine.