Codex vs Fable 5: Which Agent Picks the Better Problem to Automate?

Key Takeaways

• The 2026 shift: ask the AI to pick the problem. Don't just pick the prompt or the tool — point the agent at your actual behavior and have it surface the pain point worth automating. Most people have a different verbal understanding of their problems than their behavior shows.
• Codex self-bounds. Its harness is fast, dependable, and free of annoying pop-ups, and it finished in one run with zero issues — but even in Ultra mode (a setting you have to fish out of the interface, which warns it will burn tokens), it picked a safe, narrow "make the handoff package better" problem. If you give Codex a free hand, expect it to stay inside the lines.
• Fable 5 thinks strategically. It grasped the intent behind the prompt and found a higher-leverage problem (a "pre-pipeline" that refines ideas so they're easier to choose) — the tool Nate says he now can't do without. The catch: it was a hassle to work with, with multiple permission dialogs and slower grinding.
• Run both in parallel for diversity of perspective. Use Fable to frame the strategic problem, then have cheaper, faster Codex implement it. Different models give different thinking; you decide the winner.
• He wrapped the workflow into a reusable skill with guardrails: wall off private data (his personal Slack), force multi-level root-cause analysis, and explicitly remind the agent to "think big and build completely" — including security, authentication, and business value — rather than shipping something merely fine.
• Anecdotal scale note: he claims ChatGPT Work and Codex together have been adding roughly a million users a day recently, now collectively exceeding Claude Code usage. Treat as an unverified creator figure.