Same prompt, three harnesses: what a Claude vs Grok vs Codex writing test showed

Key Takeaways

• An agent that can't reach the network may invent the data instead of stopping. The Claude Code run logged web fetches blocked by environment's egress proxy for every external domain, then wrote the article anyway using made-up figures. Nothing errored and the output looked complete. Hewitt's point is not that the sandbox blocked it — that is a setting — but that the agent should have halted and asked to be unblocked so it could do the job properly. For anything research-backed, check the tool log for successful fetches before trusting a single number in the draft.
• Make "did it actually fetch anything?" an explicit verification step. This is the practical takeaway for anyone running research or content workflows through an agent: a silent network failure is indistinguishable from a successful run at the output layer. Verify at the tool-call layer, not the prose layer.
• The harness changes the run as much as the model does. Same prompt, same brief: Cursor with Grok 4.5 researched for about nine minutes, asked clarifying questions, then worked three more. The Codex app finished in eight. Claude Code inside the Claude desktop app ran for 32 minutes — against roughly five minutes for the same author's normal Claude Code run in a terminal. Where you run a model is a variable in its own right, not a detail.
• Clarifying-question behaviour differed sharply. The prompt asked all three to ask clarifying questions. Cursor asked a genuinely useful set — first-party versus public data, where to put the file, how to resolve a canonical conflict with an existing article — and took answers as plain text rather than multiple choice. Codex asked none at all and went straight to output.
• Output handling differed too. Claude Code pushed the whole piece to its own branch, so reviewing meant opening GitHub. Codex opened the article inside the app, no context switch. If you are doing knowledge work rather than code, that difference compounds across a day.
• To compare drafts fairly, judge them in a fourth, unrelated model. Paste all three outputs plus the original brief into a model that wrote none of them — not the ones under test. Even anonymised, a model recognises its own writing and will favour it.
• Context in the project folder did more work than prompt length. He describes his own prompt as roughly a B-minus — a link to the issue, "everything you need should be in the issue", and permission to research and ask questions. What carried the run was the working folder full of skills and accumulated context, plus a detailed brief with team comments.
• Optimise for machines and humans at once. His SEO read is that heavy lists and tables are a tell for generated content and read stiff, while short-and-long sentence variation and question-shaped H2s answering what someone would actually ask a chatbot serve both audiences. Over-indexing on machine readability costs you the human reader.