Codex's runtime in the Agents SDK alpha
OpenAI on Nov. 3 (Nov. 4 Japan time) released "v0.22.0-alpha.1," an alpha version of the "OpenAI Agents SDK," a software development kit (SDK) for JavaScript and TypeScript, as a prerelease in its official GitHub repository. The release notes say the harness of the coding tool "Codex" has been embedded in the SDK. The harness is the runtime that manages AI model calls and tool execution, and the Codex app, CLI (command-line version) and IDE (development environment) extensions also run on it.
The alpha is a trial version still under development, and the latest version on npm (a JavaScript package distribution service) remains 0.21.0. No announcement on OpenAI's blog or official X account had been found as of Nov. 5. The release was introduced in a post on Nov. 4 (Nov. 5 Japan time) by Daniel San, who leads the AI unit at hedgineer.ai, and not in an official OpenAI announcement.
They're embedding the Codex harness directly into the Agents SDK
San said the change lets developers use Codex inside the SDK instead of building and maintaining their own runtime, adding that it is still an alpha. What the release notes confirm is only that the harness takes charge of managing the processing. Whether developers will no longer need their own runtime is not stated in the release notes and remains unconfirmed.
After September's Agents API, now inside apps too
OpenAI has been working to open Codex's underlying system to outside developers. The "Agents API," which entered public beta in September, is a service in which OpenAI runs the Codex harness, and on Oct. 30 (Oct. 31 Japan time) it added features including "Auto Review," in which a separate AI reviews risky operations. The Agents SDK, by contrast, runs inside a developer's app, with deployment, data storage and approval of operations controlled on the app side. This alpha puts the Codex harness into that app-side SDK.
Since April, harness and sandbox (isolated workspace) features had been added to the Agents SDK in stages. However, OpenAI has not explained how the new "embedded Codex harness" differs from the earlier features. The "Codex SDK," which lets programs control a local Codex, continues to be offered as a separate product.
- Harness and sandbox execution added to the Python Agents SDKReport
- TypeScript v0.9.1 supports sandboxes and the harnessThird party
- Agents API public beta beginsOfficial
- Five features including Auto Review added to the Agents APIOfficial
- Codex harness embedded in the Agents SDK alphaOfficial
Harness manages history and context summarization
According to the changelog, when a developer runs an agent (a system in which AI plans steps and carries out work), the Codex harness manages the exchanges with the model, tool execution, history and context summarization. Developers can define their own tools in TypeScript and can also use tools based on MCP, a common standard for connecting external tools, as well as remote execution environments.
It also includes features for long-running work. Developers can monitor work in progress, intervene midway and cancel it, and, where the state supports it, save and restore as well as branch sessions (units of a sequence of work) that are waiting. Policies for approving tool execution can also be applied. A testing feature lets developers try out behavior with scripted responses, without a real model or API key.
Entry points change for existing apps, and state cannot be carried over
What affects existing developers is a change in import paths. In this alpha, @openai/agents becomes the entry point for the Codex harness, and the previous API has moved to @openai/agents/classic. The Realtime API for voice is also used from /classic/realtime.
Session state is separate between the previous approach and the harness approach. Converting in-progress state from the previous approach into state for the harness is not supported. OpenAI also provides @openai/agents/migration, which runs some previous definitions on the harness, but describes it as experimental.
Installation requires Node.js 24 and files distributed on GitHub
Supported environments are Linux x64, macOS on Apple silicon and Windows x64, and Node.js 24 is required. Linux also requires glibc 2.39 or later and OpenSSL 3, and Windows requires the Microsoft Visual C++ runtime. The SDK is not obtained from npm; it is installed by placing the files distributed on GitHub locally. To add it to an existing app, four packages must be matched to the same version and the same environment. The bundled samples call a real model with an OpenAI API key configured.
Pricing and general release timing unconfirmed
The release notes, changelog and installation instructions do not mention pricing or when a general release will come. They also do not say which models can be used, whether it can be used with a ChatGPT account or whether it will come to the Python version, and all of these remain unconfirmed. ORIENS could not confirm whether a feature equivalent to the Agents API's Auto Review is available, or how Codex's sandbox works.
Development is still moving. The change that added the experimental harness (#2113) asked reviewers to check points such as how credentials and saved state are handled, but the outcome could not be confirmed. On Nov. 5, a change (#2157) that splits the runtime into a separate package, "@openai/agents-codex," was also merged into the development branch, and the number of packages needed for installation is expected to go from four to five. After the merge, one test for macOS failed, and no cause specific to this change has been identified.
Anthropic and Cursor have similar SDKs
Competitors are also moving to provide agent runtimes as development kits. Anthropic's "Claude Agent SDK" makes the same tools, processing flow and context management as Claude Code available from Python and TypeScript, and the company describes it as a library that runs the Claude Code executable. AI coding tool maker Cursor began general availability of its "Agent SDK" on Oct. 26 (Oct. 27 Japan time).
