A Few Patterns We Frequently Use With Fable 5

A Few Patterns We Frequently Use With Fable 5 ![rw-book-cover](https://readwise-assets.s3.amazonaws.com/static/images/article4.6bc1851654a0.png) 🔗 **Source:** [A Few Patterns We Frequently Use With Fable 5](https://twitter.com/ClaudeDevs/status/2074606058128224365#thread) > [!note]+ Document Notes > tags: [model-routing, ai-agents, llm-cost-optimization, sub-agents, prompt-caching] > author: [MISSING] > source: X (@ClaudeDevs) > reference: [MISSING] > date created: 2026-07-09 09:14 > > # 🏷️ Title: A few patterns we frequently use with Fable 5 (thread) — @ClaudeDevs >  ## 🧭 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 - Source thread references SWE-bench Pro, BrowseComp, and Claude Managed Agents; no explicit links were included in the captured text. ## 📝 Summary - **What happened.** A short Anthropic/Claude developer thread describes two production cost-routing patterns for pairing a stronger model (Fable 5) with a cheaper executor (Sonnet 5). - **Why it matters.** Both patterns recover most of the frontier model's task performance while paying mostly the cheaper model's rate — a direct lever on agent operating cost. - **Key stakeholders.** Teams building multi-agent systems on Claude Managed Agents; anyone optimizing the cost/quality tradeoff in LLM pipelines. - **Risks and tradeoffs.** The quoted benchmark deltas (8% quality loss on SWE-bench Pro, 4% on BrowseComp) are vendor-reported single-benchmark figures, not independently reproduced, and generalization to other task types is unstated. - **Bottom line.** Route rare high-value reasoning to the expensive model and delegate token-heavy work to the cheap one; per-sub-agent caching keeps repeated context from being paid for twice. ## 📖 Main Content - **Advisor pattern.** An executor (Sonnet 5) does the bulk of the work and calls Fable 5 only occasionally for guidance. Because Fable 5 is invoked rarely (~once per task), most tokens bill at the lower executor rate. Reported result on SWE-bench Pro: ~92% of Fable 5's standalone score at ~63% of the price. - **Orchestrator pattern.** Fable 5 plans and delegates, while Sonnet 5 worker sub-agents execute the token-heavy work. Reported result on BrowseComp (Claude Managed Agents, Fable 5 orchestrator + Sonnet 5 workers): 96% of Fable 5 performance at 46% of the price. - **The unifying idea** is model-tier routing by token volume: keep the expensive model's token footprint small (either as a rarely-called advisor or as a planning orchestrator that doesn't do the grunt work), and push high-volume tokens to the cheaper tier. - **Caching detail.** Claude Managed Agents supports both patterns with sub-agents; each sub-agent keeps its own cache, so repeat calls don't re-pay full price for the same context. - **Terms.** *Advisor* = escalate up for guidance; *orchestrator* = delegate down to workers. *SWE-bench Pro* is a software-engineering task benchmark; *BrowseComp* is a web-research/browsing benchmark. - **Uncertainties.** "~once per task" is approximate; the price percentages are relative to an all-Fable-5 baseline whose absolute cost isn't given; no error bars, sample sizes, or task-mix details are provided. ### 🗒️ Notes - Two named patterns, one shared principle: minimize expensive-tier token share. - Advisor: executor-led, expensive model as a consultant. ~92% quality / ~63% cost on SWE-bench Pro. - Orchestrator: expensive model leads planning, cheap workers execute. 96% quality / 46% cost on BrowseComp. - Orchestrator shows a better cost/quality frontier in these two data points (higher quality retention *and* lower relative cost), but the benchmarks differ, so the comparison isn't apples-to-apples. - Per-sub-agent cache is the mechanism that makes repeated delegation affordable. - Direction of escalation is the mnemonic: advisor = up, workers = down. ## 📚 References - SWE-bench Pro — software engineering agent benchmark (referenced, not linked) - BrowseComp — browsing/research agent benchmark (referenced, not linked) - Claude Managed Agents — sub-agent framework supporting both patterns (referenced, not linked) ## 🗣️ Paraphrase - Let a cheaper model run the task and phone the stronger model for advice only when needed; you keep most of the strong model's quality at a fraction of the cost. - Or flip it: let the strong model plan and hand the heavy lifting to cheaper workers; on a research benchmark this held 96% of quality at under half the cost, with per-worker caching preventing duplicate context charges. ## 📰 Newsletter Not applicable to this source — the content is model-routing engineering for Claude agents, not a WordPress or community item. ## ✅ Action Items #### DevRel - Developer experience: write a short recipe/example showing the advisor vs. orchestrator wiring in Claude Managed Agents, with the token-share intuition up front. - Education: a "when to escalate up vs. delegate down" decision snippet for agent builders. (Remaining Action Items subteams — OSPO, Community, WordPress, The WPCC, FAIR — Not applicable to this source: it's an LLM-agent cost pattern, not open-source governance or WordPress project work.) ## 🧪 Hypotheses - H1: The advisor pattern's cost savings shrink as tasks require more frequent Fable 5 consultation (i.e., savings are inversely proportional to advisor-call frequency). - H2: The orchestrator pattern's quality retention degrades on tasks where planning and execution are tightly coupled (harder to cleanly delegate). - H3: Per-sub-agent caching yields larger savings on tasks with high shared-context reuse than on tasks with mostly unique context. - H4: On SWE-bench Pro, applying the orchestrator pattern would beat the advisor pattern's 92%/63% frontier. ### 📊 Methodology - Run both patterns on the *same* benchmark set (not one each) with matched task samples; report quality and cost with confidence intervals. - Sweep advisor-call frequency and worker count to map the cost/quality frontier. - Ablate caching on/off to isolate its contribution to the reported price reductions. ### ❓ Questions - What is the absolute (not relative) cost of the all-Fable-5 baseline these percentages are measured against? - How was "~once per task" advisor invocation enforced or measured? - Do the patterns compose (orchestrator whose workers themselves use an advisor)? ### 🔮 Future Research - Cross-benchmark generalization of both patterns beyond SWE-bench Pro and BrowseComp. - Automated routing policies that pick advisor vs. orchestrator per task at runtime. ### 💡 Implications - Practice: agent architecture is now a cost-engineering discipline; tier selection by token volume is a first-class design decision. - Policy/security: delegating to sub-agents widens the surface for prompt injection and context leakage across cache boundaries — worth an audit. - Education: benchmark-relative claims need absolute baselines to be actionable. ### 🧠 Learning Styles Assessment - Reading level estimate: roughly Flesch-Kincaid grade 10–12 (dense technical noun phrases, short sentences). Formula: Flesch-Kincaid Grade Level, estimated (not computed). #### 👀 Visual learners - Present: none — text-only thread, no diagrams. - Improve: a two-panel diagram (advisor: executor→advisor; orchestrator: orchestrator→workers) with cost/quality bars. #### 🔊 Aural learners - Present: none. - Improve: a 60–90s audio walkthrough of the two patterns. #### 🗣️ Verbal learners - Present: clear pattern names and definitions (advisor, orchestrator, escalate up / delegate down). - Improve: a worked example with concrete task and token counts. #### 🧩 Kinesthetic learners - Present: none. - Improve: a runnable Claude Managed Agents snippet for each pattern. #### 🧠 Logical learners - Present: quantified cost/quality tradeoffs (92%/63%, 96%/46%). - Improve: a decision tree for choosing a pattern by task shape. #### 🧑‍🤝‍🧑 Social learners - Present: thread format invites replies. - Improve: a discussion prompt on where each pattern breaks. #### 🧘‍♀️ Solitary learners - Present: self-contained, quick read. - Improve: a self-check ("which pattern minimizes expensive-tier tokens here?"). #### 🌿 Naturalistic learners - Not applicable to this source — no environmental/ecosystem dimension. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: plain paragraph text, no images needing alt. - Improve: if republished, use real headings/landmarks rather than tweet blocks. #### Keyboard navigation - Not applicable to this source — static text, no interactive UI captured. #### Contrast and readability - Not applicable to this source — rendering is platform-controlled; no authored color. #### Text resizing and reflow - Not applicable to this source — no authored layout. #### Media and captions - Not applicable to this source — no audio/video/media in the captured text. #### Forms and interactions - Not applicable to this source — no forms. #### Standards reference - If diagrams are added later, provide text alternatives per WCAG 2.2 AA 1.1.1. ### 🌍 DEIB Assessment #### Cultural breadth - Not applicable to this source — narrow technical scope, no cultural content. #### Inclusive language - Present: neutral technical language; jargon (advisor, orchestrator, sub-agent) mostly defined in context. - Improve: expand SWE-bench Pro / BrowseComp on first use for non-specialists. #### Neurodiversity support - Present: short, chunked, consistent structure. - Improve: a one-line TL;DR before the two patterns. #### Representation - Not applicable to this source — no people or identities depicted. #### Psychological safety - Not applicable to this source — neutral technical framing, no interpersonal content. ### ✅ Logical Validity #### Strong reasoning - Causal/mechanistic: the claim that cost drops because "most tokens are billed at the lower executor rate" is a sound mechanism directly tied to the pattern design. Location: advisor and orchestrator paragraphs. - Inductive: the two benchmark results support the general principle of tier-routing by token volume — reasonable but based on two data points. #### Fallacies - Hasty generalization (mild): "patterns we frequently use" plus two single-benchmark numbers implies broad applicability that the evidence doesn't establish. Fix: state task scope and add cross-benchmark results. - Missing baseline / relative-figure framing: percentages are relative to an unstated absolute cost. Fix: report absolute costs alongside percentages. #### Tone of voice - Confident, practitioner, matter-of-fact. Example: "A few patterns we frequently use with Fable 5" — presents patterns as established practice. ### 🛡️ DARVO Analysis - Not applicable to this source — no accusation, defense, or conflict; it's a neutral technical thread. ## 🎯 Conclusion - Bottom line: route expensive-tier tokens sparingly. Use Fable 5 as an occasional advisor to a Sonnet 5 executor, or as an orchestrator delegating to Sonnet 5 workers; both keep most of the quality at a large fraction of the cost, and per-sub-agent caching avoids paying twice for shared context. - Next steps for readers: try the pattern that matches your task shape (advisor for execution-heavy tasks, orchestrator for research/planning-heavy ones), and validate the cost/quality tradeoff on your own workload rather than trusting the benchmark deltas. ## 🔄 Refactor Appendix — A few patterns we frequently use with Fable 5 > 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] Advisor pattern bills most tokens at executor rate > [!quote] ID:rw_01kx4nkgbehd6mvphy3d6sj5q0 > Use Fable 5 as an "advisor." An executor (Sonnet 5) calls Fable 5 for guidance. Most tokens are billed at the lower executor rate. **Context & use**: - Opening pattern in the thread. - Core cost mechanism: rare expensive-model calls keep the token bill on the cheap tier. **Tags**: #xref **Highlight**: [view highlight](readwise://read/01kx4nkgbehd6mvphy3d6sj5q0)  #### [H2] Advisor holds ~92% quality at ~63% cost > [!quote] ID:rw_01kx4nkgbehd6mvphy3d6sj5q0 > Sonnet 5 + a Fable 5 advisor tool gets ~92% of Fable 5's score at ~63% of the price. **Context & use**: - SWE-bench Pro result. - Quantifies the advisor tradeoff; note it's a single-benchmark, vendor-reported figure. **Tags**: #xref **Highlight**: [view highlight](readwise://read/01kx4nkgbehd6mvphy3d6sj5q0)  #### [H3] Orchestrator holds 96% quality at 46% cost > [!quote] ID:rw_01kx4nkgbehd6mvphy3d6sj5q0 > The Fable 5 orchestrator achieves 96% of Fable 5 performance at 46% of the price. **Context & use**: - BrowseComp result with Sonnet 5 worker sub-agents. - Better cost/quality point than advisor here, but on a different benchmark. **Tags**: #xref **Highlight**: [view highlight](readwise://read/01kx4nkgbehd6mvphy3d6sj5q0)  #### [H4] Per-sub-agent cache avoids double-paying for context > [!quote] ID:rw_01kx4nkgbehd6mvphy3d6sj5q0 > Each sub-agent keeps its own cache, so repeat calls don't pay in full for the same context twice. **Context & use**: - Closing detail on Claude Managed Agents. - Explains why repeated delegation stays cheap. **Tags**: #xref **Highlight**: [view highlight](readwise://read/01kx4nkgbehd6mvphy3d6sj5q0)  ### 🧠 Permanent Candidates (H4) #### [P1] Route expensive-tier tokens by volume, not by task **Claim** (one sentence, Courtney's voice): Multi-model agents get cheaper without losing much quality when the expensive model handles only low-volume reasoning and the cheap model handles high-volume execution. **Evidence pointer**: [[#H1]] **Scope & conditions**: Holds when tasks separate cleanly into rare high-value reasoning and abundant routine tokens; breaks when the hard reasoning is itself token-heavy or inseparable from execution. **Links**: - [[Multi-agent architecture]] — because applies-in agent design - [[LLM cost optimization]] — because extends  ### ✏️ Fleeting Queue - [F1] Compare advisor vs. orchestrator on the *same* benchmark to see which frontier wins — #fleeting #inbox - [F2] Find the absolute cost baseline behind the "% of price" figures — #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Draft a one-page advisor-vs-orchestrator decision note for agent builders — DevRel #things  📄 **Full text:** [[Full Document Contents/Articles/A Few Patterns We Frequently Use With Fable 5.md|A Few Patterns We Frequently Use With Fable 5 by @ClaudeDevs]] ## Highlights - Use Fable 5 as an "advisor." An executor (Sonnet 5) calls Fable 5 for guidance. Most tokens are billed at the lower executor rate. ([View Highlight](https://read.readwise.io/read/01kx4nsnj63qrz5j6sh88p9cfm)) - Note: The advisor pattern + its cost mechanism in one line. - Tags: atom/literature/h1 - Sonnet 5 + a Fable 5 advisor tool gets ~92% of Fable 5's score at ~63% of the price. ([View Highlight](https://read.readwise.io/read/01kx4nsw7v5tngxscr5zwjahgb)) - Note: Quantifies the advisor tradeoff on SWE-bench Pro. - Tags: atom/literature/h2 - A second strategy: use Fable 5 as an orchestrator. Fable 5 plans and delegates to workers (Sonnet 5). ([View Highlight](https://read.readwise.io/read/01kx4nt4tktp9c023a25fxqd13)) - Note: Defines the orchestrator pattern. - Tags: atom/literature/h2 - The Fable 5 orchestrator achieves 96% of Fable 5 performance at 46% of the price. ([View Highlight](https://read.readwise.io/read/01kx4ntbef2g4rdvj6bqvg1ms2)) - Note: Quantifies the orchestrator tradeoff on BrowseComp. - Tags: atom/literature/h3 - Each sub-agent keeps its own cache, so repeat calls don't pay in full for the same context twice. ([View Highlight](https://read.readwise.io/read/01kx4ntmjs8jr3x0s6n7q1f5p8)) - Note: The caching mechanism that keeps delegation affordable. - Tags: atom/literature/h4