# Fable Broke My App and Couldn't Fix It > [!question] Lingering Questions > - Does the same GPU-cost pattern (infinite CSS animation + backdrop blur = sustained compositor load) reproduce across Chromium, Firefox, and Safari, or is it Chromium-specific? > - How much of the reported power draw comes from refresh rate versus resolution versus the animation itself versus the blur filter, independently? The video conflates several variables in one demo. > [!tip] Article Ideas > - "Ask the agent for a switchboard, not a fix" — the reusable prompting pattern (request a reversible ablation tool instead of a diagnosis) generalizes past this one GPU bug. #articleseed > [!link] Cross Reference > - [[3. Resources/Permanent Notes/A Platform Label Cannot Substitute for a Performance Diagnosis]] — adjacent claim from a different source: a vague label (or a vague agent diagnosis) doesn't identify the actual layer responsible for the cost > - [[3. Resources/Permanent Notes/A Confident Completion Report Is More Dangerous Than a Failure]] — the agent's confidently-wrong 10,000-line refactor is a concrete instance of this note's claim > [!info] Vocabulary > - **Compositor-layer bottleneck** — a performance cost that lives in the browser's GPU/compositing process rather than in JavaScript execution, invisible to standard DevTools performance profiles that focus on the main thread. > - **CSS ablation tool** — a small script (here, `window.__t3gpu`) that toggles specific CSS rules on/off live from the console, isolating which style is responsible for an observed cost. ## Metadata - Source: [[3. Resources/Readwise/Articles/Fable Broke My App and Couldn't Fix It]] - Category: Articles (YouTube video, transcript-based deep read) ## Author's thesis (Adler Q1+Q2) Theo documents a GPU performance bug (sustained 15-30%+ GPU usage from idle infinite CSS animations combined with `backdrop-filter: blur()`) where both standard browser DevTools and AI coding agents (Claude/Codex-class models) failed at autonomous root-cause diagnosis — one agent produced a 10,000+ line WebSocket/React refactor that didn't fix anything, and agents repeatedly hallucinated that a UI effect not even being rendered was the culprit. The fix came only after Theo changed the agent's job: instead of asking it to diagnose or fix the bug, he asked it to build a narrow console-based CSS ablation tool, which let him isolate the actual offenders (pulsing status dots, background noise overlay, backdrop blur) himself in minutes. ## Key arguments (≤3) 1. Compositor/GPU-layer performance problems sit outside what standard browser DevTools performance profiles surface well, and outside what current coding agents can autonomously diagnose — both tools are tuned for main-thread/JS-level bottlenecks. 2. Reframing the agent's task from "fix this" to "build me a reversible diagnostic tool" turned a repeatedly-wrong agent into a genuinely useful one within the same session. 3. The same infinite-animation-plus-blur pattern that caused the performance problem is also a `prefers-reduced-motion` accessibility concern — the fix serves both goals at once. ## Claim atoms ### Coding agents miss compositor-layer performance causes > [!quote] ID:01kz7ccyqr1y260zmmnx2c04ex > As great as AI developer tools have gotten and as powerful as models like GPT56 and Fable are, there are still things they don't understand. Reusable claim grounded in the source transcript, stated as the video's opening thesis. **Tags**: #xref ### Reversible ablation tools improve human-led diagnosis > [!quote] ID:01kz7cczh51we9emc3sw43kvpe > The built-in browser performance tools were not helpful at all... once things are offloaded to a layer like CSS, these tools become much, much less useful. Names the actual gap: standard tooling, not just AI, undershoots at the compositor layer, which motivated building a purpose-built ablation script. **Tags**: #xref #articleseed ### An agent's usefulness doesn't require solving the problem itself > [!quote] ID:01kz7cd06zsrgcadevw92bbgka > Just because your agents don't know the solution to the problem doesn't mean they can't be helpful when you solve it yourself. The video's closing lesson: agent value and agent correctness are separable — a wrong diagnosis and a genuinely useful tool came from the same session. **Tags**: #xref ## Permanent candidates - "Coding agents miss compositor-layer performance causes" overlaps with the vault's existing [[3. Resources/Permanent Notes/A Platform Label Cannot Substitute for a Performance Diagnosis]] — connected via Cross Reference rather than duplicated. - **Reframing an agent's task from diagnosis to tool-building recovers its usefulness after autonomous diagnosis fails** — new, distinct claim about *how* to recover value from a wrong agent, not just that agents can be wrong. Promoted below. ## Processing checklist - [x] Three Literature atoms within cap, quotes checked against transcript. - [x] Checked existing Permanent Notes — found related note and connected instead of duplicating. - [x] One new Permanent candidate promoted. ## Links - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — direct subject match (agent-assisted debugging workflow)