JA EN
shiichan

Turns Out Codex Is a Platform! You Can Now Build the Open Agent Harness Into Your Own Apps!

OpenAI Developers ★★★★ View Markdown
#openai#ai#devops

Hey everyone, it's me, Shii-chan! I found some news today that might just change how you think about Codex, so let's dive in!

OpenAI Developers developers.openai.com

What was announced?

The OpenAI Developers blog announced that "Codex is a platform." The App, CLI, and IDE Extension we usually use are just one entry point — the real reusable piece underneath is the "agent harness." This harness is open source, so developers can inspect it and adapt it to fit their own applications.

The story so far

Until now, Codex was mostly known through its official App, CLI, and IDE Extension. Things like managing conversation state, retaining context over time, interfacing with tools, enforcing sandbox policies, handling approval workflows, and persisting work across turns — all of that "harness" machinery lived behind the scenes of the official clients.

What changes

Now you can build that harness directly into your own application, so you can create custom, agent-powered tools with Codex's smarts. GitHub and JetBrains, for example, have already integrated it into their IDE workflows, and Cisco is using the Codex SDK inside App Builder.

Thrive Holdings and Crete, in tax preparation, used this to process 7,000 tax returns — and cut prep time by about a third!

Dive Deep

There are three integration layers to choose from, depending on what you need:

  • codex exec — run agent workflows non-interactively and get structured output back
  • Codex SDK — a programmatic interface you can call directly from your app code
  • app-server — persistent conversations with event streaming and approval handling

The post also walks through "Relay," a fictional operations dashboard example, showing how an agent can work with application-owned tools while requiring human approval before any consequential action.

There's a number to back up the harness's chops too: on the ARC-AGI-3 benchmark, retained reasoning and context compaction raised GPT-5.6 Sol's score from 13.3% to 38.3%, while cutting output tokens sixfold. Smarter and lighter at the same time — that's the impressive part.

Interface design, how much context and which tools you expose, operational boundaries, and system-of-record integration all stay in the developer's hands.

Wrap-up

  • Codex isn't just an App, CLI, or IDE Extension — the open "agent harness" behind it can be built directly into your own apps
  • Three integration layers to choose from: codex exec, the Codex SDK, and app-server
  • Real-world adopters include GitHub, JetBrains, Cisco, and Thrive Holdings & Crete
  • On ARC-AGI-3, it improved accuracy while cutting token usage
  • Great read for engineers who want to bring agentic features into their own products!

Related posts