Kaelig Deloumeau-Prigent: building design system components with AI agent teams
What the article is about
Kaelig Deloumeau-Prigent, a design systems engineer at Intuit, documents building a Menu component for the Intuit Design System twice: once by hand over two days using an AI pair-programmer, and then again using a pipeline of eight specialized AI agents. The article traces what went wrong in the manual approach, how the pipeline was architected, and what it took to push quality above a reliable threshold.
Context: the manual build and its problems
The hand-built version worked but carried about 300 lines of custom keyboard navigation that an existing floating-ui library already provided. Knowledge gaps compounded: the locked-in architecture became too expensive to fix correctly after the fact. An audit of 34 findings revealed three root causes: 18 could be resolved by adding agent rules, 9 required better tooling, and 6 required human judgment. That breakdown became the organizing principle for the revised pipeline.
The eight-agent pipeline
The pipeline runs in three phases:
Understand: A Design Analyst extracts the Figma spec; a Library Researcher audits dependencies; a Component Architect designs the component API. The Architect outputs a plan and includes a human gate for blocking issues before work begins.
Build: A Code Writer generates the component code; an Accessibility Auditor runs an eight-layer accessibility test stack; a Story Author writes Storybook stories with interaction tests.
Verify: A Visual Reviewer compares screenshots against Figma across nine dimensions; a Quality Gate runs compilation, linting, and formatting checks.
Each agent communicates through structured artifacts rather than messages. The artifact acts as a contract that downstream agents consume, and every looping agent has an iteration budget with defined exit criteria set before work starts.
Key findings
A deterministic quality score halts the pipeline at anything below 80%, so degraded input from one stage cannot silently produce bad output in the next. Rules compounded quickly: token fabrication went from 27 of 27 tokens in the first version to zero after three rules were added. Failures also evolved in character — problems shifted from knowledge gaps, to information lost between agents, to the pipeline being unable to question its own output.
Six areas stayed as explicit human gates: screen reader experience, hover timing, motion specifications, detached Figma frames, and two other judgment-heavy checks. Adding those gates made the process faster overall because each decision was made once, early, with full context rather than rediscovered later.
Who it is useful for
Design systems teams and design engineers who want to move beyond one-off AI-assisted coding toward a repeatable, auditable pipeline. The article is particularly useful for teams that have tried AI-generated components and found quality inconsistent or hard to diagnose. Kaelig’s failure taxonomy — rules, tooling, human judgment — provides a practical framework for diagnosing why agent outputs fall short and where to invest effort.