OpenAI has made the Agents API, which provides developers with the open-source harness and infrastructure that powers Codex, available as a public beta. The API enables production-ready agents to be created with a single call by specifying the task, model, tools, and runtime environment. OpenAI hosts and maintains the harness, while developers can choose the sandbox managed by OpenAI, their own infrastructure, or environments provided by partners.
Integrations for the Agents API are available with Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. Introduced for those who want to get started quickly, the OpenAI-hosted sandbox uses the sandbox infrastructure used in Codex and ChatGPT. In this environment, provided and managed by OpenAI, agents can run code, work with files, and produce outputs. The sandbox can be configured with files, packages, capabilities, and extensions.
The API provides versioned access to capabilities alongside model versions and continuously improves the harness together with the models. In long-running sessions, context management helps the agent continue working by automatically compressing previous information when the session approaches its context limit. This means developers can reportedly create workflows spanning multiple context windows without building their own context-compression mechanisms.
The tool search feature aims to reduce token usage and costs by loading the definitions of the tools needed when required. Programmatic tool calling allows agents to perform operations in parallel, chain related operations, and filter and combine results with code. The Agents API supports built-in tools such as MCP, custom functions, and web search.
The multi-agent architecture breaks complex tasks into independent parts and distributes them to sub-agents in parallel. While each sub-agent maintains its own context, the main agent coordinates the results. This feature can be used in research, analysis, and coding processes.
The Agents API is built on the open-source Codex harness, which manages model calls, tools, and context. While OpenAI operates and maintains the system, developers can inspect its source code. The service is available to all developers during the beta period at no additional cost; only the tokens and tools used are billed.
Why it matters
This expansion aims to ensure that developers do not have to build operational components such as the runtime environment, tool management, and context preservation for long-running tasks from scratch in agent applications. Different sandbox options and infrastructure integrations allow applications to run either in an environment managed by OpenAI or on infrastructure chosen by the developer. While the open-source nature of the Codex harness enables examination of how model calls, tools, and context are managed, keeping the responsibility for operation and maintenance with OpenAI separates the ways in which the system can be used. Although the lack of additional charges during the beta period makes access easier, total costs will be determined by the tokens and tools used; how costs will take shape, particularly in long sessions, parallel operations, and multi-agent workflows, remains an open question.
Background
The API is not a new name in the FikirPilot archive: we published a news article mentioning this name in the last 90 days; that article was dated 13 September 2026.
Term: agent
An artificial intelligence agent is software that calls tools and carries out multi-step tasks to achieve a goal instead of producing a single response.