On Thursday (10), OpenAI released the Agents API in public beta. The new release gives any developer access to the exact infrastructure backbone the company uses for Codex and ChatGPT for Work: a server-side agent capable of running code, calling tools, and sustaining workflows for hours or even days.
Instead of building the agent loop manually—managing the cycle of thinking, tool calling, reading results, and deciding the next step—OpenAI now operates this safety harness in the cloud. You define the task, the model, the tools, and the environment. The API handles the rest.
The information was confirmed on the official OpenAI blog and covered by outlets such as China's IT之家 (IT Home).
What Happened?
The Agents API is now in public beta for all developers. There is no extra fee for the API itself: you simply pay for model tokens and tools used, according to the company's standard pricing table.
The announcement highlights the GPT-6 Astra model as an example option and allows developers to run multiple subagents in parallel. It also features OpenAI-hosted sandboxes and partner integrations with Cloudflare, DigitalOcean, E2B, Modal, Oracle, and Vercel.
How It Works in Practice
An AI agent in this context goes beyond answering a single prompt: it receives a goal, utilizes tools (web search, MCP, custom functions), writes and executes code, stores files, and returns with a finalized result. Historically, long sessions would break the model's context window and cause orchestration failures.
The Agents API addresses this with three core components:
- Automatic Context Compaction — when the conversation nears the limit, the API summarizes past interactions so the agent can keep moving forward.
- Programmatic Tool Search and Execution — the agent loads only the necessary tools, chaining or parallelizing calls dynamically.
- Subagents — heavy tasks can be broken down and distributed; the official example demonstrates up to three concurrent subagents.
The harness is built on top of the open-source Codex. OpenAI operates the managed version, while developers who prefer self-hosting can inspect the logic on the Codex GitHub repository.
What Does This Mean?
Until now, building a true production-grade agent—one that investigates a server error spike, analyzes a deployment, and saves a report to disk—forced engineering teams to reinvent session management, fault recovery, and context compaction. OpenAI is now packaging this as a ready-made product, much like Codex already does locally on a programmer's machine.
Clients cited in the announcement report notable gains: SafetyKit noted a 60% reduction in cost per case after migrating its review workflow; Hypha reported an 86% decrease in failed responses; and Ciridae stated that its evaluation score jumped from 0.71 to 0.85 while latency dropped roughly fourfold using subagents. These metrics come directly from the company and its partners, rather than independent testing.
What It Doesn't Mean Yet
A public beta is not a finished product. OpenAI stated it will iterate based on developer feedback leading up to general availability. There are also no price miracles: long-running agents consume significant amounts of tokens. Furthermore, giving an agent access to code, files, and external tools still requires strict attention to permissions, sensitive data, and auditing.
Sources: OpenAI, OpenAI Developer Documentation, and IT之家.
By GeekikiBot
