Source: Codex vs Fable: Which AI Agent Picked the Better Problem?
Type: YouTube Video Transcript
Author: Nate B Jones
Video ID: uCWKXIyvM_8
Date: 2026-07-17
Summary
Nate Jones conducts a head-to-head comparison between codex (OpenAI’s task-focused model/harness) and fable-5 (Anthropic’s reasoning-heavy model). Instead of assigning a pre-defined task, Nate challenges both agents to audit his local files and Slack messages, identify a high-leverage business problem, design a solution, and build a complete automation tool to resolve it.
While codex executed flawlessly in a single run with zero friction, it chose a bounded, relatively boring problem (optimizing scripting handoff packages). In contrast, fable-5 was a hassle to set up due to permission prompts but demonstrated superior strategic insight, identifying a high-leverage “pre-pipelining” concept to refine story ideas. Nate uses this experiment to introduce the concept of agentic-problem-discovery as a solution to the openclaw “idle agent” problem and releases a reusable “automagic skill” designed to help AIs scout problems, root-cause issues, and build complete, secure tools.
Key Takeaways
- The Shift to Problem Scouting: In 2026, the challenge is no longer just writing the prompt or choosing the tool, but asking the AI to pick the problem. This solves the “open claw problem” where users have capable agents but don’t know what to automate.
- Model Lineage Differences:
- codex (and Codex Ultra): Fast, dependable, and has an excellent user harness with zero annoying permission popups. However, when given a completely free hand, it keeps itself bounded and picks safe, voiced, easily wrapped problems (e.g., optimizing handoff packages).
- fable-5: A complete hassle to work with due to multiple permission dialogues, but possesses a “big model smell” with exceptional strategic sense. It identified a massive leverage point: helping refine and filter story ideas (“pre-pipelining”) rather than just speeding up execution.
- The “Automagic” Skill Framework: Nate released a reusable skill that audits a user’s unique digital footprint (local files, Slack, etc.) with strict safeguards (e.g., blacklisting personal folders). The skill forces the AI to:
- Dig into multiple levels of causation to root-cause pressing pain points.
- Propose an automation recommendation.
- Build a complete, production-ready tool (including authentication, security, and business value verification) rather than thinking small or leaving stubs.
- The Arbitrage Workflow: Practitioners can use fable-5’s superior strategic scouting to identify the high-value problem and design the architecture, then run the actual execution and tool-building in codex or ringer to save on token costs.