
OpenAI's Gabriel Chua published a useful explainer on how to think about Codex — their agentic coding system — by decomposing it into three layers: the model (intelligence), the harness (execution environment), and the surfaces (where you interact with it). As a mental model for understanding what "Codex" actually means when people say "Codex," it's the clearest thing I've read on the subject.
The distinction between model-alone and model-plus-harness — "without the harness, you have suggestions; with the harness, you have execution" — is clarifying. That's the line between autocomplete and something qualitatively different, and most coverage of agentic coding tools muddles it completely.
But here's the move worth noticing: the post presents this three-layer decomposition as though it's how agentic coding works, not how OpenAI's product works. There's no mention that Claude Code, Cursor, Jules, or anyone else exists — let alone that they make fundamentally different architectural choices. The "mental model" is shaped exactly like the product it's describing. Funny how that works.
The athlete-and-racquet analogy is where the quiet part gets loud. The model and harness are "co-designed," Chua writes — you wouldn't swap an athlete's racquet before a competition. Read: don't expect our model to perform as well in someone else's tooling, or someone else's model to perform as well in ours. That's a lock-in argument wearing an engineering costume. It might even be true. But plenty of teams are building harness-agnostic workflows specifically because they've learned what happens when you let a single vendor own every layer of your stack.
The closing line — that Codex can help you "ship complete work" — arrives without a single qualifier. No failure rates. No task-complexity caveats. No mention of the supervision overhead that "agentic" still demands in practice. For a post aimed at developers evaluating this for their teams, that's not an oversight. It's a choice.
Good explainer. Just remember who drew the map.
Stay current weekly
Get new commentary and weekly AI updates in the AI Primer Briefing.