OpenAI Launches Public Beta of Codex-Based Agents API

On September 10, 2026, OpenAI launched the Agents API in public beta, built on the harness and infrastructure of Codex. Leveraging long-running agent management experience accumulated from operating Codex and ChatGPT for Work, this API enables the creation of production-ready agents with a single call.

Key Features and Architecture

• Harness Management: OpenAI hosts and maintains the harness, with compute environments selectable from OpenAI-managed sandboxes, self-hosted infrastructure, or partner sandboxes.
• Multi-Agent Support: Complex tasks are split into independent parts and delegated in parallel to subagents. Each subagent maintains its own context, while the main agent handles coordination and result integration.
• Efficiency Optimization: Tool search loads only necessary tool definitions to reduce token usage and costs, while Programmatic tool calling supports parallel calls and result filtering.
• Long Session Support: When approaching context limits, previous contexts are automatically compressed (compacting) to preserve information, eliminating the need for developers to implement their own compression logic.

Partner Ecosystem and Performance

Integrated with various sandbox providers such as Blaxel, Cloudflare, Daytona, and E2B, offering flexible VPC deployment and resource configuration. Early users have reported significant performance improvements.

• Ciridae: Evaluation scores rose from 0.71 to 0.85, with latency reduced by 4x.
• SafetyKit: Cost per case decreased by 60%, with improved token efficiency.
• Hypha: Failed agent responses decreased by 86% through the separation of harness and sandbox.

Costs and Future Plans

During the current public beta period, there are no usage fees for the Agents API; only token and tool costs incurred by agents are billed. OpenAI is preparing for General Availability (GA) based on user feedback.