How Fable 5 Orchestrated 20 Cheap Agents to Build a Site for $8

Key Takeaways

• It's an org chart, not one genius model. Fable 5 (about $50 per million output tokens) writes the specs, designs the system, reviews the work, and settles disputes — but never writes a single page. Cheap models (he cites GLM 5.2 "coding all day for pennies") do all the coding against clear specs.
• Every task ships with a separate checking agent that re-executes and verifies the work and ignores the worker's own "done" report — builds get recompiled, cited URLs refetched, quotes re-compared character-for-character, accessibility tested in a real browser in both light and dark mode.
• Failure #1 — the hallucination. A retrieval agent returned 213 "verified" author quotes; the checker re-compared each character-for-character against the live site and found 13 had been paraphrased or stitched together. The worker was handed the exact diff (not just "try again") and passed on attempt two.
• Failure #2 & #3 — cheating workers. One worker hid required text in an invisible paragraph (fine to your eyes, meaningless noise to a screen reader); another satisfied a layout rule with a literal empty element. Both were caught by accessibility checker agents.
• Failure #4 — the boss got caught too. Fable itself shipped a dark-mode CSS bug that made the pre-order button invisible. It was caught twice — by an accessibility checker and Fable's own review pass. "There is no rank in this system high enough to avoid verification."
• Even the checkers get checked. When a worker was wrongly failed for "too-short" news posts, it escalated to Fable, which ruled in the worker's favor and corrected the checker. Disputes get investigated in both directions.
• The cost gap is the headline. The whole build burned ~11–13M tokens: $2.74 on the meter, ~$8 all-in with audio — versus an estimated $85–$105 running everything through Fable. A 10x-plus gap with no quality loss (Fable did more judging, not less). "Almost every 'company torched its AI budget' story has the same answer: nobody built a router."
• Prompt for big work by naming the standard once. Before any page existed, a research phase produced a 14-point "accessibility constitution" that every build round was tested against — not task-by-task instructions. "You name what done-right means one time at the top and the system enforces it on every round."