+ STUDY:
DESIGN AGENT LOOPS
ED. 01 2026

BREAK THE
WEBSITE

STUDY: DESIGN AGENT LOOPS, EDITION 1 · LIVE CANVAS, SELF-HEALING · JULY 2026

AUTONOMOUS IMPROVEMENT LOOP · WITH DEPLOYMENT GUARDRAILS

THIS PAGE IMPROVES ITSELF

This loop runs continuously on the site's own usage data and was designed to observe and learn from which surfaces break most. Based on live user data and breakage, the agent forms a hypothesis, puts each proposed change through deployment guardrails, ships only what passes, and measures its performance.

LIVEREAL, FROM USAGE DATA MODELEDFORWARD PROJECTION
MODELED GENERATION 001
MODELED CHANGES SHIPPED 0
MODELED REGRESSIONS PREVENTED 0
MODELED MEAN TIME-TO-RESTORE 14.8s
LIVE

LIVE DEMAND: AWAITING LIVE DATA…

MOST-REQUESTED REPAIR: AWAITING LIVE DATA

GEN 001
  1. 1OBSERVE
  2. 2HYPOTHESIZE
  3. 3GATE
  4. 4DEPLOY
  5. 5MEASURE

↻ ONE GENERATION ≈ ONE DAY OF TRAFFIC, COMPRESSED FOR THE DEMO

STANDING BY

MODELED DEPLOYMENT GUARDRAIL STACK

EVALS
VISUAL-REGRESSION DIFF ·
CONTRAST / WCAG AA ·
LAYOUT REFLOW · NO OVERLAP ·
PERFORMANCE BUDGET (LCP, BUNDLE) ·
POLICY
CHANGE CONFINED TO ALLOWLISTED SURFACE (#HERO, BUTTONS) ·
WITHIN CHANGE BUDGET ·
APPROVAL
RISK SCORE ·
ROLLOUT
CANARY 10% → 50% → 100%, HEALTH-GATED ·
REVERSIBILITY
PRE-CHANGE SNAPSHOT CAPTURED · ROLLBACK ARMED ·

AWAITING FIRST CYCLE.

MODELED PERFORMANCE DASHBOARD

MEAN TIME-TO-RESTORE
FIRST-PASS FIX RATE

GEN 001 BASELINE → CURRENT: MTTR +0.0S, FIX-RATE +0.0PTS

IMPROVEMENT LOG

    ABOUT

    AUTHOR'S NOTE

    Study: Design Agent Loops is an independent project built to educate through illustration, showing how agent loops work by way of a familiar surface: design and website maintenance. It is one approach to demonstrating agent systems working live. Critically, the website was built entirely by agents, and it uses a set of parallel agent systems to keep improving itself.

    To illustrate and explore these self-healing loops, the website is divided into two parts. The first is a hero section that asks you to break the website in one of three ways, injecting a regression on purpose. An agent is then deployed to the live runtime, notices the damage, and works through the steps a careful human would: diagnosing the cause, proposing a fix, creating checks, and ultimately restoring the website to its original form.

    The second part looks at the data infrastructure that the agent both maintains and references, using it to learn more about the breakages visitors trigger and the restoration methods applied.