the comic experiment

If you give AI agents millions of tokens, what decides whether the comic is any good? Three rounds, three complete graphic novels, one honest human judge between each. Same theme every time: ordinary people waking up to the first, faint superpowers. Everything else changed. Read all three and watch what actually moved the quality.

Afterlight, page 1: an eclipse over five cities round 1 · baseline

afterlight

An eclipse wakes abilities on five continents. 50 agents each drew one page as raw SVG code. No image model, no editor, no second look.

  • one agent wrote all 50 page beats in a single pass
  • 50 agents drew their page as hand-written vector code
  • no critique loop, no visual review

Judge's verdict: consistent, coordinated, and soulless. And too close to its inspiration.

50 pages51 agents3.4M tokens$0 images~30 min
read it
The Warm Hill, page 1: a Silesian mining town at dusk round 2 · the medium fix

the warm hill

A dying Silesian mining town, a caregiver who absorbs pain, a son who hears metal break. The art switched from code to a painting model, and a critic entered the loop.

  • writer plus one merciless critique pass, then revision
  • 50 agents each directed an image-model page and checked it with their own eyes
  • strict character tokens held the look across 50 isolated generations

Judge's verdict: the style is much better, but the story is quiet to a fault. Little dialogue, no action.

50 pages53 agents3.0M tokens$9 images~34 min
read it
Small Hours cover: a sleeping face dissolving into a night map of the world round 3 · the judgment fix

small hours

Five countries, five faint first mutations, a dream-finder stitching them together, and a containment company that is not entirely wrong. Written from a detailed human brief.

  • a real brief from the human client, encoded as hard requirements
  • two critique passes: structure and character, then dialogue and action
  • 65 art directors with visual QC, then editor agents re-read every chapter as drawn and ordered redraws

Judge's verdict: much better. The first one that reads like a comic.

65 pages151 agents9.4M tokens$24 images~2 hrs
read it

what the experiment found

Token count barely mattered: rounds 1 and 2 cost the same and could not be further apart in quality. What moved the needle each time was structural: matching the medium to the tool, putting critics and reviewers between draft and delivery, and a human's taste steering the brief.

Agents scale logistics. Taste needs judgment loops.