OpenAI expands Responses API into full agent platform with shell, sandbox and context compaction

OpenAI has made a major move for developers building AI agents. On March 27, 2026, the company launched a massive expansion of the Responses API, turning it into a complete platform for autonomous agents.

The key additions include a built-in shell tool that lets agents run command-line operations directly, a hosted container sandbox where agents can safely work with files and network access, and an agent execution loop that allows the model to propose and execute actions iteratively until the task is complete.

Context compaction addresses one of the biggest challenges in long-running agent tasks. The system compresses previous steps into shorter representations, allowing agents to work across many iterations without hitting token limits.

OpenAI has also introduced reusable agent skills for defining complex, repeatable tasks. The company recommends the Responses API for all new projects and plans to formally deprecate the Assistants API by mid-2026.

The expansion is a clear response to competition from Anthropic, which has gained significant traction with Claude Code and agentic tools. OpenAI is now positioning itself to reclaim the developer market with a more integrated and powerful platform.