OpenAI Releases Agents API in Public Beta

OpenAI has released its Agents API in public beta, granting developers access to the same harness and infrastructure that powers Codex. This managed service is designed to support the creation of long-running AI agents, which require robust context management, efficient tool utilization, and reliable infrastructure capable of operating for extended periods, potentially days. The announcement highlights that the Agents API is now live and available for all developers.

The Agents API is built upon the open-source Codex harness. OpenAI's internal experience scaling Codex and ChatGPT for Work informed the development of this API, addressing the specific needs of long-running agents. These needs include a sophisticated harness for managing context, effective coordination of sub-agents, and efficient tool usage. The accompanying infrastructure is engineered for sustained reliability. Developers can choose to run their agent's compute within an OpenAI-managed sandbox, their own infrastructure, or a partner sandbox. Data processed through the API remains US-only, and Zero Data Retention is not supported.

The API's structure is organized around four core concepts: Agent, Environment, Session, and Events/Items. An Agent is defined by its model, instructions, available tools, and Multi-Agent Coordination Protocol (MCP) servers. An Environment is an optional sandbox where the agent can access files, load skills, and execute commands. A Session represents a persistent agent instance dedicated to a task, responding to input. Events and Items refer to the inputs sent to the agent and the outputs it generates. The process of running a session involves four steps: creation with a defined task, progress monitoring via streaming or webhooks, and the ability to continue with new tasks or steer the current turn.

OpenAI provided a practical example of the API's capabilities by demonstrating the creation of an incident-investigation agent with a single API call. This example showcases the integration of the `gpt-6-astra` model, the use of MCP servers for observability via HTTP, and the enablement of multi-agent coordination with a maximum of three concurrent sub-agents. The agent is configured to access tools and capabilities within a specified environment, including a partner sandbox. The `vault_ids` parameter suggests integration with a secure storage system for sensitive information, such as `vault_YOUR_VAULT_ID`.

The public beta launch signifies a significant step in making advanced AI agent development more accessible. By abstracting away much of the underlying infrastructure complexity, OpenAI aims to empower developers to build more sophisticated and autonomous AI applications. The focus on managed services and robust infrastructure addresses common challenges in deploying and maintaining AI agents, potentially accelerating innovation in areas like automated customer support, complex data analysis, and personalized digital assistants.