Codex: Not for Songs, Built for Long Agentic Work

Codex Isn't Slow, It's Playing a Different Game

Here Codex appears not broken, but tuned for tasks where it's better to double-check an action than silently wreak havoc. OpenAI's documentation describes an auto-review mode: the model passes the planned action and fresh context to an approval sub-agent, which can auto-approve low-risk steps and some riskier ones with sufficient authorization.

On a creative task this turns into bureaucracy. A user asks for a song and gets five clarifying questions, a concept, a writing plan, and approval requests after each step. Formally, the model makes the process more meticulous, but the result doesn't materialize where a quick aesthetic leap is needed.

And here the contrast is more interesting than the complaint. Other discussion participants write that Codex runs for half a day or longer, and one user describes a week of autonomous work and thousands of experiments. I wouldn't tout this as a verified benchmark: it's user experience, not a reproducible test. But the direction aligns with how OpenAI positions Codex: safe defaults, an isolated environment, long agentic cycles, and work with code and research tasks.

So the problem isn't that the model can't handle text. The problem is its behavior is optimized for planning, actions, verification, and risk control. For a song, it sounds like pedantry. For experiments, it already resembles useful discipline.

Where This Model Character Truly Shines

The main takeaway is simple: evaluate Codex not as a universal author, but as an agentic system for tasks with verifiable steps. Coding, research, hypothesis testing, experiment launches, and iterative refinement benefit far more from this caution than they suffer.

In open-ended writing the opposite is true. There often isn't a single success criterion, but there is style, pace, bold phrasing, and readiness to deliver a cohesive draft right away. If the model instead builds a protocol of approvals, it breaks the flow. Not because it's stupid, but because it tries to be reliable where taste is expected.

For me, the important part of this story isn't arguing which model writes better songs. The input data doesn't give a fair ranking between Gemini, Claude, Grok, Fable, or others. But it clearly shows the engineering fork: the same caution can be annoying friction in text and a valuable safeguard in autonomous work.

The most dangerous conclusion would be to call Codex a bad model. A more accurate one is duller but more useful: agentic behavior without the right script quickly looks like bureaucracy with a GPU.

Previously, we covered how Anthropic reversed hidden query underranking in Claude, restoring transparency after the scandal. This change caused the model's overcaution, leading it to refuse song generation while remaining useful only for scientific tasks.