Codex vs Fable: Which AI Agent Picked the Better Problem?

Nate B Jones · 2026-07-17

This video presents a fascinating experiment comparing the AI agents Codex 5.6 and Fable in solving an open-ended problem: automating a business process. The core challenge was to have the AIs not just execute a task, but first discover the problem by analyzing real-world data (Slack, local files) and then proposing an automation solution. Fable demonstrated superior strategic problem identification, finding a more impactful and leveraged problem, while Codex opted for a more bounded, shape-focused approach. The video concludes by introducing a reusable Automation Discovery and Build skill designed to empower users to leverage AI for problem identification and robust automation creation across various platforms like Codex and Fable, emphasizing the importance of strategic problem-finding over tool selection.

read more

Nate B Jones presents a captivating comparison between Codex 5.6 (representing OpenAI’s agent capabilities) and Fable (an alternative AI agent), specifically focusing on their ability to solve an open-ended business automation challenge. The central premise of the experiment, termed "Find the job, then build it," was to task both AIs with discovering a recurring job within Nate's existing business processes and then automating it. This was a novel approach, as Nate explicitly instructed the AIs not to ask him which workflow to automate, but to identify three credible opportunities, choose one, and then build a controlled, safe, and repeatable automation without incurring additional costs or making unapproved external mutations.

The findings revealed a stark contrast in the AIs' strategic thinking and problem identification. Fable, despite being more challenging to set up due to multiple permission dialogues, ultimately excelled at strategic problem identification. After grinding through the data, Fable identified a problem that was much more interesting and leveraged. It recognized that one of the hardest jobs in the storytelling business (Nate's domain) is finding the right story to tell, i.e., determining what matters to talk about in a world of infinite AI stories. Fable proposed a "Slate Preflight" tool, a pre-greenlight duplicate and coverage checker for candidate topics. This tool surfaces nearest coverage and pipeline collisions, records deterministic results, and offers safe refusals, providing a powerful start gate for the content creation workflow. This demonstrated Fable's ability to understand the intent behind the prompt and discover a problem with significant strategic value.

In contrast, Codex (specifically in its "Ultra" setting, which allowed it to burn a tremendous number of tokens and gain extensive context from Nate's Slack and files) took a more narrow and shape-focused approach. Codex identified the problem of ensuring Content Package Handoff Preflight, a deterministic readiness check for completed content packages. This involved checking package shape and readiness, but not semantic meaning. While Codex successfully built a functional automation that was quick and efficient (one run, zero issues), Nate noted it was a typically Codex-flavored definition of the problem. It focused on the form of the problem rather than its deeper strategic implication. This highlights a key "con with Codex": when given an open-ended problem space, it tends to keep itself bounded, even with immense computational resources. It picks problems it can easily "wrap its arms around," focusing on readiness rather than semantic alignment.

Nate emphasizes that this is not a benchmark to declare one AI superior overall, but a case study revealing their differing strengths. Fable's strategic thinking is crucial for identifying high-leverage opportunities, guarding the start of the workflow. Codex, with its efficiency and dependability, is excellent for executing bounded tasks, acting as a handoff gate. The ideal scenario involves using both: Fable for discovery and strategic problem identification, and then potentially Codex for efficient and cheaper execution of the resulting automation.

To help others leverage these insights, Nate developed a reusable skill called "Automation Discovery and Build." This skill is designed to: 1) Map available work surfaces, 2) Observe repeated human glue (hidden work), 3) Qualify three candidates for automation, and 4) Build and prove one winner. The key takeaway is to "Ask the AI to find the job before naming the tool." This approach empowers users to uncover their own hidden automation backlogs—the "repeated human glue" appearing across chat (Slack), tracking (Linear handoffs), and artifacts (email and docs rework). By feeding their unique data fingerprints to an AI, users can gain a tailored automation solution, effectively creating a "magic easy button" for their business. This skill also incorporates safeguards, allowing users to define boundaries (like excluding personal Slack channels), ensuring responsible and controlled AI use. The ultimate goal is to leverage AI's non-deterministic nature to launch effective, automagic skills that audit unique user fingerprints, building neatly tuned automation solutions for the most pressing problems, while using cheaper execution engines (like Ringr, which Nate mentions as cheaper than Fable) for the final implementation.

Smartest #model should not be writing most of your #code #ai #agenticengineering #fable

Agentic Engineering · 2026-07-17

This video proposes a layered AI model architecture where a powerful, expensive model acts as an architect/planner and a reviewer, while cheaper, faster models serve as builders. The architect model's role is to understand the codebase, identify high-impact problems, and generate precise, self-contained implementation instructions with clear success criteria. Subsequently, the builder models execute these detailed tasks, and the architect returns to review the results against the original goals. This approach aims to optimize both cost and efficiency by leveraging the strengths of different AI models for distinct cognitive tasks.

GLM 5.2 launch boosts coding AI with 1M token memory, local use #GLM5 #AIlaunch #OpenSourceAI

AI Honeycove · 2026-07-16

The GLM-5.2 open-source model, developed by China, offers capabilities surpassing Claude Code in several areas, including an impressive 1-million-token context window and persistent memory for extended multi-step coding projects. It can be run entirely locally, eliminating API token costs. The installation is straightforward, requiring a single terminal command and API key setup.

Codex Was A Developer Brand, So Why Remove It?

Theo - t3.gg · 2026-07-14

This video argues that OpenAI's decision to rebrand Codex as ChatGPT and then Claude Code was a significant miscalculation. The speaker posits that Codex had cultivated a strong, loyal developer community and brand identity, independently of OpenAI's other offerings. By deprecating the Codex brand, OpenAI destroyed that allegiance and unique appeal, losing an organic, developer-centric connection that neither ChatGPT nor Claude Code could replicate.

Architects, Anti-Patterns, and Organizational Fuckery

Charity Majors · 2023-03-09 · 27 min read

TLDR: The "architect" role becomes an anti-pattern when it severs decision-making authority from the people who bear the consequences of those decisions — the engineers who build and operate the systems. Architecture should be a core skill expected of senior engineers, not a separate function held by someone who's drifted away from the codebase; the QA analogy is instructive: outsourcing the skill weakened everyone. The real organizational dysfunction is using brilliant engineers as glorified political middlemen instead of fixing the structural problems that make such mediation feel necessary.

Deploys Are The ✨WRONG✨ Way To Change User Experience

Charity Majors · 2023-03-08 · 10 min read

TLDR: Deploys and releases are distinct operations that most teams wrongly conflate — deploys should happen continuously as a pure engineering concern, while user-facing changes should be controlled independently via feature flags. This decoupling lets you flip a feature at a precise moment (like a 10am press embargo) without touching the deploy pipeline, and enables progressive rollout to subsets of users so you can safely validate behavior in production before full exposure.

June 2026 newsletter

Simon Willison · 2026-07-03 · 1 min read

This is a paywall announcement for Simon Willison's June 2026 monthly newsletter covering the current state of LLM tooling, including notable model releases like Claude Fable 5, GPT-5.6, and GLM-5.2, plus updates to his own open-source tools like sqlite-utils 4.0 and Datasette. It matters primarily as a signal that the LLM landscape is moving fast enough that a respected practitioner sees value in a monthly digest to track meaningful shifts in models and tooling.

Nano Banana 2 Lite

Simon Willison · 2026-06-30 · 1 min read

Google released Gemini 3.1 Flash Lite Image, a fast and cheap image generation model accessible via the API as "gemini-3.1-flash-lite-image." For engineers evaluating text-to-image models for cost-sensitive, high-throughput applications, this represents a new low-cost tier in the Gemini image model lineup, though quality limitations like inconsistent text rendering are still present.

A few questions about what you're working on...

Kent Beck · 2026-03-04 · 2 min read

Kent Beck is running a reader survey to better understand who his audience is and what problems they're facing, so he can make his writing on software design more relevant. For a senior engineer, the only signal worth noting is in the comments, where a 20-year veteran flags a real tension: AI tools are accelerating delivery but may be preventing junior developers from building the judgment needed to distinguish good solutions from bad ones.