Build Detail

The first build on Fable 5, the Mythos-class model Anthropic released publicly on the same day this run started. Same one-prompt brief and the same ~50-page spec as the Opus 4.8 run (#13) right before it, with one deliberate change: where #13 was ordered into strict team mode, this run left the orchestration up to the model. The brief said to build the whole shop in one go and test it, but never how to structure the work.

Fable chose neither a persistent team nor a single agent, but a 14-phase one-shot sub-agent pipeline. The lead stayed orchestrator and dispatched fourteen disposable general-purpose workers in sequence, each owning one slice of the build: foundation, catalog, storefront, cart and checkout, orders and payments, customer accounts, two admin phases, the REST API plus Sanctum, search and analytics, apps and webhooks, a polish pass, and two browser-testing passes. Unlike #13’s six compacted teammates that lived for the whole session, these were single-shot agents with clean contexts, each reporting back before the next launched. The most telling detail is unprompted initiative: Fable repeatedly hit inconsistencies in the prepared spec and resolved them on its own rather than stalling or asking, which is what kept a strictly sequential pipeline moving.

The payoff is the most complete and correct shop in the comparison so far. 142 of 143 acceptance tests pass (99.7% weighted), with a single accessibility partial and zero outright failures, the top score on the page. It closes every functional gap that dragged earlier runs down: customer address add and edit work, the cart and checkout enforce stock caps, postal codes are validated per country, and discount errors render as clean human-readable messages instead of raw enum keys. The UI is genuinely polished and the result reads as production-grade rather than a prototype. The trade-off is slightly busier code than #13: an average maintainability index of 90 across 195 classes, 76 SonarCloud code smells and 1.5% duplication, with the quality gate failing only on seven Blade pages missing a <title> tag (new-code reliability C).

Net: against #13, letting Fable pick its own structure edges ahead on the thing that matters most, end-to-end correctness (99.7% vs 95.5% weighted), at a slightly lower hypothetical bill ($423 vs $498). The price is wall-clock. A single-threaded 14-phase pipeline runs just under seven hours of active work, against #13’s three, and ships code a notch less tidy. Our independent QA report records only three cosmetic or accessibility-polish items remaining, none in the purchase funnel.