From Models to Agents: Building Computer Environments into the Responses API

A shift is underway from models that simply perform specific tasks to agents capable of handling complex workflows. However, building agents brings practical challenges related to setting up execution environments, such as intermediate file storage, network security, and timeout handling.

To address this, OpenAI has combined a shell tool with a hosted container workspace in the Responses API. When a model proposes a step, the platform executes it in an isolated environment, carrying out real work through a file system and restricted network access.

The new shell tool offers a much broader range of use than the existing Code Interpreter (Python-only). It supports Unix utilities such as grep, curl, and awk by default, and enables complex agentic tasks like running Go or Java programs or starting a NodeJS server.

The Responses API handles orchestration between the model and hosted tools. When the model proposes a shell command, the API service passes it to the container runtime, and by repeating a loop that feeds output results back to the model in real time, it completes the task.