OpenAI has expanded its Responses API by adding a computer environment, marking a shift from static models to dynamic agents capable of handling complex workflows. This new capability allows models to execute real-world tasks like running services, requesting data from APIs, and generating artifacts such as spreadsheets or reports.

Building agents traditionally creates practical challenges: managing intermediate files, handling large data tables, providing network access without security risks, and implementing timeouts and retries. Instead of requiring developers to build their own execution environments, OpenAI has integrated these components directly into the Responses API, combining it with a shell tool and hosted container workspace.

The shell tool dramatically expands what models can accomplish by enabling command-line interactions through familiar Unix utilities like grep, curl, and awk. Unlike the previous code interpreter limited to Python, this tool supports multiple programming languages including Go, Java, and NodeJS. The system orchestrates an execution loop where the model proposes actions, the platform runs them in isolated environments with restricted network access, and results feed back into the workflow.