OpenAI launches Agents API to handle agent orchestration

OpenAI has released its Agents API in public beta, turning the orchestration harness behind Codex and ChatGPT for Work into a fully managed service. The API coordinates model calls, context management, and subagents on OpenAI's infrastructure, eliminating the need for developers to build their own custom orchestration layers. The service is available with no additional fees beyond standard token and tool consumption, and the underlying harness remains open-source through the Codex repository.

The system includes built-in features designed to manage long-running sessions. Automatic context compaction summarizes earlier interactions to prevent sessions from hitting context limits, while tool search loads definitions on demand to save tokens. Programmatic tool calling allows agents to execute parallel requests, and a simple configuration flag enables parallel subagents to divide complex tasks. For execution, developers can use an OpenAI-hosted sandbox or deploy across partner environments including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel.

Early adopters of the API report significant performance improvements. The startup Ciridae saw its evaluation score rise from 0.71 to 0.85 alongside a 4x latency reduction on subagent workflows. SafetyKit achieved a 60 percent reduction in cost per case for its review workflows, and Hypha recorded an 86 percent drop in failed agent responses by decoupling the harness from the execution sandbox.

The tool layer natively supports Model Context Protocol servers, custom functions, and built-in capabilities like web search. By absorbing the operational burden of retry logic, context window management, and sandbox lifecycles, OpenAI aims to streamline agent development. While this locks the orchestration layer to OpenAI's ecosystem, it allows teams to focus their resources on proprietary tools and data rather than maintaining complex backend loops.