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

OpenAI is shifting its focus from models designed for specific tasks to agents capable of handling complex workflows. The Responses API is being equipped with a computer environment, addressing challenges like file management, network access, and timeout handling. This environment utilizes a shell tool and hosted container workspace, allowing models to propose and execute commands in an isolated setting. Key features include concurrent execution, output capping for context efficiency, native compaction for long-running tasks, and secure network access via a sidecar egress proxy. Reusable 'skills' are also introduced, packaging common workflows for consistency and efficiency. This new system allows for more robust, repeatable, and safer production workflows, moving beyond the limitations of solely relying on model prompting.

(Source：OpenAI)