Blog

5 min readfeature

The encore round: an agent that critiqued its own design work, then beat its own gates

The gallery's eighth app, built inside a measured design loop — every screenshot scored, every failure ledgered, every number traced to a receipt.

  • #landing
  • #demo

The first seven gallery apps asked whether the agent could carry a whole SDLC. This round asked the question that decides whether any of it scales: can the agent tell that its own work is not good enough yet — with numbers, not vibes — and then fix what the numbers say?

Short answer: the loop shipped. It found 23 font sizes and 32.7% off-grid spacing on a page six rounds of polish had produced, caught a 42px mobile overflow its own writing sessions had introduced, caught a contrast fix that broke the light theme one iteration after the fix landed, and drove its own flagship to every objective design gate green. The gallery now frames eight agent-built apps.

The instrument came first

The round began by building the thing the platform lacked: a critique lane that produces receipts, not opinions. scripts/design_critic.mjs runs against any route of the production build — per-band and full-page screenshots at desktop and mobile, an objective in-page audit (WCAG contrast on every visible text node against its composited background, a type-scale census, 4px-grid conformance, heading order, alts, overflow), cold-load FCP/LCP/CLS/TBT, and a strict-JSON seven-axis VLM critique through the Agnes lane — aggregated into one committed receipt per run.

A judge that cannot discriminate is worthless, so the lane ships with a known-bad calibration fixture — eleven ad-hoc type sizes, arbitrary odd spacing, sub-AA greys, an h1→h4 skip. The fixture scores 1.5; the real landing page scores 6.4. The judge discriminates.

Then the first receipt landed, and it was uncomfortable: the landing page — six rounds of agent-built polish — resolves to 23 distinct font sizes (the elite baseline is about seven) and 32.7% off-grid spacing. The deliberately-chaotic calibration fixture uses ten sizes. Our real page was less disciplined than a page designed to be bad.

The flagship

Encore — the earned performance — is the gallery's eighth app, live at /encore: a showcase, not a SaaS domain, whose content is this round's own measured story. The baseline receipt card (the honest 6.40/10 with its failing gates). The critique loop as an interactive stepper. The tool cards. The failure ledger. And the centerpiece: a before/after drag comparison between the landing page's baseline hero and the flagship's final hero — the program's success question rendered as a slider.

It was built by the agent across 17 ledgered sessions (459 turns, 513 tool calls, zero human edits to app code), phase by phase from its own 192-line build contract — the design system pinned before a line of UI existed: an eight-size type ramp, the 4/8px spacing grid, both schemes WCAG-AA, the 390px covenant, motion 150/250/400ms behind prefers-reduced-motion.

The encore hero — the axis-score strip animating the baseline means

The loop, in receipts

Every iteration of the flagship was critiqued by the same instrument, and the receipts are committed:

Run VLM overall Font sizes Off-grid Contrast (dark / light) Overflow
the landing baseline 6.40 23 32.7% 0 / unaudited 0px
iter0 — first build 6.27 6 0% 0 / 0 42px
iter1 6.00 6 0% 0 / 0 42px
iter2 6.00 6 0% 0 / 27 0px
iter3 — final 6.22 6 0% 0 / 0 0px

The trajectory of the OBJECTIVE gates is the story: the discipline the baseline lacked (type scale, spacing grid) held from the very first build — because the plan enforced it — while the loop caught and fixed what the writing sessions could not see: the 42px mobile overflow (an sr-only data table whose layout box the scoped table CSS re-expanded past the viewport), and the round's favorite failure — a contrast fix that pinned its colors "under any scheme" and broke the light theme for 27 text pairings, caught by the both-scheme audit one iteration later, fixed by inheriting the landing's own dual-declared tokens.

The iterations band — the before/after comparison and the delta table

And the honest part, on the record: the VLM overall (6.22) did not clear the 7.0 floor, and sits within a single judge's variance band. The objective gates are laws and they pass; the aesthetic floor is a goal and it stays open. The showcase says exactly that — no number without its receipt, no spin on the ones that hurt.

What the platform keeps

The round's failures are ledgered (F14–F18 in the validation log), and each one left a permanent artifact behind:

  • the critique lane with its both-scheme audit (the judge's own light-scheme blind spot was the fourteenth failure it caught — its first run photographed only the light theme and nobody noticed until the receipts described "off-white backgrounds" on a dark-first page);
  • the calibration fixture — the judge's discrimination pinned by a receipt, not asserted;
  • the seed drift gate — the showcase's every number tied to its committed receipt by a test that reads the JSON and compares field-by-field;
  • the project memory AutoDream consolidated from the build's own session notes — the sibling-inspection rule, the happy-dom origin quirk — already rendered into .unown/memory/project.md for the next round's sessions to inherit.

The primary success question — would a restarted project, using the improved platform, create a significantly better website than it could have at the beginning? — now has a measured answer. At the beginning: 23 font sizes, 32.7% off-grid, no instrument to see either. After: a first build that lands at 6 sizes and 0% off-grid because the contract and the critique loop exist, with every failure it introduces caught by a receipt, not a reviewer.

That is the encore: the performance after the performance. The next one starts from these receipts.

Spotted something wrong or worth adding? The record is the repo — open an issue.

All posts