OpenAI is extending its Responses API to provide a built-in foundation for building autonomous agents. New capabilities include a shell tool (supporting Unix utilities and languages beyond Python like Go, Java, Node.js), a built-in agent execution loop, a hosted containerized workspace with managed network access, context compaction for long-running tasks, and reusable 'skills' as composable building blocks. The execution loop lets the model propose actions that are run in a controlled environment with results fed back iteratively. Credentials remain hidden from the model via placeholders, and context compaction prevents token limit issues over extended runs.
Sort: