Ask Fable to Make Codex the Workhorse  🔗 **Source:** [Ask Fable to Make Codex the Workhorse](https://twitter.com/steipete/status/2074638582418231495#thread) > [!note]+ Document Notes > tags: [ai-coding-agents, model-routing, codex, developer-tools, llm-cost-optimization] > author: steipete (Peter Steinberger) > source: X (Twitter) > reference: https://github.com/steipete/agent-scripts/blob/main/skills/codex-first/SKILL.md > date created: {{date:YYYY-MM-DD}} {{time:HH:mm}} > > # 🏷️ Title: Ask Fable to make codex the workhorse (codex-first) — @steipete > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [codex-first SKILL.md](https://github.com/steipete/agent-scripts/blob/main/skills/codex-first/SKILL.md) — the actual substance the tweet points to. ## 📝 Summary - **What happened.** A short tweet advising that when running a particular agent workflow, you tell Fable (Fable 5, the planner model) to route the bulk of coding execution to Codex. It links to an external `codex-first` SKILL.md that encodes the pattern. - **Why it matters.** It's a concrete instance of a broader model-routing pattern: keep expensive judgment on a capable planner, push token-heavy grunt work to a cheaper executor. - **Key stakeholders.** Developers running multi-model agent stacks (Fable as router/planner; Codex, GPT-5.5, Composer 2.5 as executors). - **Risks and tradeoffs.** Router quality gates the whole thing — a bad complexity judgment sends hard work to a weak executor or wastes the expensive model on trivia. Cost claims in replies (SWE-bench ~46% cheaper per task) are community-reported, not verified here. - **Bottom line.** Planner-plans, workhorse-executes. The depth lives in the linked SKILL.md, not the tweet body. ## 📖 Main Content - **Core argument.** In an agentic coding setup, separate two roles: a *planner* that reasons about the task and makes judgment calls, and a *workhorse* that does the mechanical bulk of code generation. Assign the planner to a capable (expensive) model and the workhorse to a cheaper executor. Here specifically: Fable 5 plans; Codex executes. - **Terms.** - *Fable 5* — used here as the capable planner/router model. - *Codex* — the code-execution model cast as the cheap "workhorse." - *codex-first* — the name of the skill/pattern; make Codex the default executor for implementation work. - *Composer 2.5 / GPT-5.5* — other executors community members route to by task complexity. - **Contested / unverified claims.** The "SWE-bench ~46% reduced cost per task" and the multi-executor routing figures come from community replies, not from the author or a cited benchmark. Treat as anecdotal until reproduced. - **Honest scope note.** This is a pointer tweet. Its own body is one instruction plus a link; the operational detail (when to route, how the skill is structured) is in the external SKILL.md, which is not included in this document. ### 🗒️ Notes - The instruction is conditional: "*If you run this workflow*" — it presupposes you're already using steipete's agent-scripts setup. - The value proposition is cost/throughput arbitrage: expensive tokens spent on planning, cheap tokens spent on the large volume of implementation output. - The author explicitly flags that the tweet is a pointer, not the substance — the SKILL.md is the real artifact. - Replies extend the idea from a fixed pairing (Fable→Codex) to dynamic routing (Fable → Codex / GPT-5.5 / Composer 2.5 by task complexity). ## 📚 References - [codex-first SKILL.md](https://github.com/steipete/agent-scripts/blob/main/skills/codex-first/SKILL.md) ## 🗣️ Paraphrase - If you're running this agent workflow, have the planner model (Fable) hand off the heavy coding to Codex as the cheap executor. - The pattern keeps planning and judgment on the expensive model and sends token-heavy implementation to a cheaper one; the details live in an external skill file, and community replies claim meaningful cost savings and multi-executor routing by task difficulty. ## 📰 Newsletter - Not applicable to this source — no WordPress angle; this is an AI-agent tooling tip. ## ✅ Action Items - Not applicable to this source — no DevRel/OSPO/Community/WordPress program actions follow from a one-line routing tip. The only concrete follow-up is personal (read the linked SKILL.md), captured in the Fleeting Queue. ## 🧪 Hypotheses - Routing bulk implementation to Codex while planning on Fable lowers per-task cost without a measurable quality drop on well-specified tasks. - Dynamic routing by task complexity (across Codex/GPT-5.5/Composer 2.5) outperforms a fixed planner→executor pairing on cost-per-solved-task. - The savings hold only when the planner's task decomposition is accurate; router error erases the benefit. ### 📊 Methodology - Run a fixed benchmark (e.g., SWE-bench Verified) three ways: single strong model end-to-end; Fable-plans/Codex-executes; and Fable dynamic-routing. Compare solved-rate and cost-per-solved-task with confidence intervals. ### ❓ Questions - Where's the crossover — at what task complexity does routing to the cheaper executor start hurting solve rate? - How does the planner signal complexity, and how often does it misjudge? ### 🔮 Future Research - Reproducible cost/quality numbers for planner+workhorse splits versus single-model baselines, since the tweet's figures are community anecdote. ### 💡 Implications - Supports a "mixture of models" operating model for coding agents: judgment on expensive models, volume on cheap ones — a cost lever for anyone running agents at scale. ### 🧠 Learning Styles Assessment - Not applicable to this source — a two-sentence tweet plus a link doesn't support a reading-level or multi-modal learning analysis. Reading level: N/A (too short to score meaningfully). ### 📊 Accessibility Assessment - Not applicable to this source — no rendered document, media, forms, or navigation to evaluate; it's plain tweet text pointing to a Markdown file. ### 🌍 DEIB Assessment - Not applicable to this source — technical routing tip with no representational, cultural, or language-inclusion surface to assess. ### ✅ Logical Validity #### Strong reasoning - Sound cost-allocation logic: spending scarce/expensive capacity on judgment and abundant/cheap capacity on volume is a coherent division-of-labor argument. #### Fallacies - Appeal to unverified numbers: the ~46% cost-reduction figure is cited from replies without a linked benchmark run. Weakness is provenance, not arithmetic. Fix: link the SWE-bench result. #### Tone of voice - Terse, practitioner-to-practitioner, prescriptive ("ask Fable to make codex the workhorse"). ### 🛡️ DARVO Analysis - Not applicable to this source — no accusation, defense, or conflict; nothing to analyze for deny/attack/reverse. ## 🎯 Conclusion - Bottom line: plan on the expensive model, execute on the cheap one — Fable plans, Codex is the workhorse. The tweet is a signpost; the real content is the linked `codex-first` SKILL.md. - Next step for readers: open the SKILL.md and decide whether the routing rules fit your own agent stack before adopting the pattern. ## 🔄 Refactor Appendix — Ask Fable to make codex the workhorse > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). > Defaults: 3 Literature, 1 Permanent. Resonance filter: if you can't > articulate in one line why a claim matters, drop it. ### 📚 Literature Atoms (H3) #### [H1] Route bulk coding execution to a cheaper workhorse model > [!quote] ID:rw_01kx4nmew84ck7s5w5y3engcx0 > a capable planner (Fable 5) delegates the bulk of coding execution to Codex as the "workhorse," **Context & use**: - Tweet body, main pattern statement. - Names the core division of labor: planner reasons, workhorse executes. **Tags**: #xref **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### [H2] Keep judgment on the expensive model, volume on the cheap one > [!quote] ID:rw_01kx4nmew84ck7s5w5y3engcx0 > keeping planning/judgment on the expensive model and routing token-heavy implementation to the cheaper executor **Context & use**: - Tweet body, rationale clause. - This is the cost-arbitrage principle generalizable beyond Fable/Codex. **Tags**: #xref #aphorisms **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### [H3] The tweet's substance is the external skill, not its body > [!quote] ID:rw_01kx4nmew84ck7s5w5y3engcx0 > the substance of this tweet is the external codex-first SKILL.md, not the tweet body itself **Context & use**: - Author's own note at the end of the tweet. - Reading discipline flag: chase the linked artifact for real depth. **Tags**: #xref **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Split agentic coding into an expensive planner and a cheap executor **Claim** (one sentence, Courtney's voice): In a multi-model coding agent, you spend expensive model capacity on planning and judgment and cheap capacity on token-heavy execution. **Evidence pointer**: [[#H2]] **Scope & conditions**: Holds when the planner's task decomposition is accurate; breaks when routing misjudges complexity and sends hard work to a weak executor. **Links** (≥1 with annotation): - [[Model routing]] — because extends - [[Mixture of experts]] — because contrasts on {static role-split vs. learned gating} - [[AI coding agent cost control]] — because applies-in {agent ops} <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Read the codex-first SKILL.md to capture the actual routing rules — https://github.com/steipete/agent-scripts/blob/main/skills/codex-first/SKILL.md #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Read the codex-first SKILL.md and note whether the routing rules fit my agent stack — #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Ask Fable to Make Codex the Workhorse.md|Ask Fable to Make Codex the Workhorse by @steipete (Peter Steinberger)]] ## Highlights - a capable planner (Fable 5) delegates the bulk of coding execution to Codex as the "workhorse," ([View Highlight](https://read.readwise.io/read/01kx4nves9tanexjjcsv2y1mra)) - Note: The planner-plans / workhorse-executes division of labor — the tweet's point. - Tags: atom/literature/h1 - keeping planning/judgment on the expensive model and routing token-heavy implementation to the cheaper executor ([View Highlight](https://read.readwise.io/read/01kx4nvn5bznzd0931gj6rj9ww)) - Note: The generalizable cost-arbitrage principle: judgment on expensive models, volume on cheap ones. - Tags: atom/literature/h2 ## Links - [[3. Resources/Permanent Notes/Platform Friction Sets the Real AI Productivity Gain]] — Routing work across models matters only when coordination overhead stays below the saved effort. - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — A tactical multi-model workflow reference.