News · 2026-08-24

OpenAI open-sourced the agent loop, not the model

OpenAI has open-sourced the Codex harness under Apache-2.0 -- the execution runtime that sits between an application and the model, managing threads, tool calls, streaming, sandboxing, and human approvals. The company published the design in a developer post by Nicolas Bonamy and Derrick Choi on August 19, 2026, and on August 24 its release notes deprecated the older codex mcp-server command in favor of the new app server. What is open is the loop; the models are not.

Key facts

• The openai/codex repository is Apache-2.0 licensed and sits at roughly 117,000 GitHub stars.
• The announcement post, Codex as a platform, is dated August 19, 2026.
• On August 24, 2026, OpenAI's release notes deprecated codex mcp-server in favor of the app server.
• Three integration layers ship: codex exec, the TypeScript and Python SDKs, and the app server.

For two years the interesting question about AI coding tools was which model was behind them. This release is an argument that the question is now obsolete, and it comes from the company with the most to lose if it is true.

Here is the distinction that matters. A model produces tokens. Getting useful work out of one requires a lot of machinery around it: something that holds context across many turns, decides which tools to expose, runs them in a sandbox, streams partial progress to a user, catches failures and retries, pauses to ask a human before doing something irreversible, and hands back a result the calling system can store. OpenAI's post names that machinery directly. "That surrounding execution system is the harness," Bonamy and Choi write. And on the boundary of what was actually released: "The open-source layer is the harness and integration surface; model access and managed services remain separate."

The technically interesting piece is the app server. It is a bidirectional JSON-RPC interface over stdio, with websocket and Unix-socket options, that streams turn/started, item/*, and turn/completed events while routing approval requests back to the calling application. In plain terms: your software can embed a live, stateful Codex agent, watch it work in real time, interrupt it mid-thought, hand it tools you wrote, and require a human click before it does anything consequential. The SDKs sit one layer up -- the Python SDK exposes thread_start, thread_resume, thread_fork, and streaming, and defaults new threads to an automatic review approval mode -- while codex exec covers bounded, non-interactive runs with structured output.

The analogy is the web browser. For a decade, the browser was the product; then the rendering engine got factored out, and suddenly every desktop app could embed a browser without building one. Codex app-server is doing that to the agent loop. OpenAI's own demo, Relay, is a fictional shipment-operations dashboard with an agent living beside it -- the product owns the records and the business rules, the harness owns the loop. "The most interesting opportunity is not to reproduce the Codex app with a different logo," the post argues, "but to build software that reflects how a specific person or team already works."

Why it matters is competitive, not technical. If the harness is the reusable part and it is Apache-licensed, the moat moves. It stops being "which model writes better code" and becomes "whose runtime is embedded in the most software." That is a durable position in a way that a benchmark lead is not, and OpenAI is not alone in noticing: DeepSeek shipped a harness where the agent loop itself is a plugin, and the community proved the point from the other direction with a proxy that keeps Claude Code's harness and swaps the model underneath. Ground Truth's explainer on agent harnesses and scaffolding covers why this layer decides so much of an agent's real behavior.

External adoption is not just marketing. Cisco says its App Builder inside Cisco Cloud Control is powered by the Codex SDK, letting users build applications in natural language inside Cisco's own platform. OpenAI also reports that Thrive Holdings and Crete ran a Codex-driven tax workflow through 7,000 returns, cutting preparation time by roughly a third.

The honest caveat is the one skeptics keep raising, and it is fair: this is wrapper-level leverage. Nothing here makes a model smarter. If the harness turns out to be mediocre, the release is packaging, and the several thousand teams who have already written their own agent loop will keep theirs. There is also a strategic reading worth holding in mind -- open-sourcing the layer that makes your model easy to embed is a good way to make your model the default thing embedded. Apache-2.0 is genuinely permissive, but the post is explicit that the managed services are not part of the giveaway.

For the underlying concepts, see AI agents and tool use and function calling.

Primary source, verified: read the paper →

Ground Truth. (2026, August 24). OpenAI open-sourced the agent loop, not the model. Ground Truth. https://groundtruth.day/news/openai-open-sourced-the-agent-loop-not-the-model.html

@misc{groundtruth:openai-open-sourced-the-agent-loop-not-the-model,
  title  = {OpenAI open-sourced the agent loop, not the model},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/openai-open-sourced-the-agent-loop-not-the-model.html}
}

Topics: agents · openai · open-source · developer-tools · harness · mcp