OpenAI Launches Agents API with Full SDK and Tooling
Original: OpenAI Agents API
Why This Matters
A unified agent platform with sandboxes, MCP, and multi-agent orchestration raises the bar for enterprise AI deployment.
OpenAI has published official documentation for its Agents API, covering architecture, session management, multi-agent workflows, sandboxes, tools integration (including MCP), observability, and an Agents SDK — signaling a major push toward production-grade autonomous agent deployment.
OpenAI's Agents API documentation reveals a comprehensive platform for building, running, and orchestrating AI agents at scale. The API centers on a Responses API with conversation state management, background mode, streaming via WebSocket, and mid-turn steering — giving developers granular control over agent behavior during execution.
Key infrastructure includes both OpenAI-hosted and self-hosted sandboxes with defined lifecycle and security controls, a file and artifact system, and deep tool integration: web search, function calling, MCP (Model Context Protocol) connections, plugins, code interpreter, shell access, and computer use. A 'Vaults' feature suggests secure credential or data storage for agents.
The multi-agent section supports orchestration across agent networks, with webhooks for async workflows and compaction for managing long context. An Agents SDK accompanies the API, offering quickstart guides, guardrails, result/state management, and integration hooks. A ChatKit layer adds UI widget customization for agent-facing interfaces.
The documentation also references GPT-6 Astra under key concepts — the first public acknowledgment of that model name in developer-facing materials — alongside reinforcement fine-tuning, evals, and a prompt optimizer. Pricing and a model catalog are included, pointing to a full productization push.