# The 4,300 prototypes problem

> What happens after the agents start shipping, and why governance is the second half of machine-readable.

Published 2026-08-07.

Here's a number that should sit on the desk of everyone selling AI-built UI, including me: one team put their winning machine-readable format into production, generated 4,300 prototypes in four months, and an audit of a sample still found typography violations, broken spacing, and an invented color palette nobody approved.

Machine-readable context, working as designed, at real scale. And the drift still got in.

I keep this number close because it's the honest counterweight to everything else I argue. I run an enterprise design system and ship agent-built software to about 8,000 users, so I have every incentive to tell you structured context solves the problem. It doesn't. It solves the first half. The 4,300 figure is what the second half looks like when it's missing.

Read the number carefully, because it gets abused in both directions. The pessimist's misread: "AI-generated UI doesn't work." Wrong. 4,300 prototypes in four months is a two-orders-of-magnitude change in exploration capacity; teams used to ship a dozen prototypes a quarter and call it a good quarter. The optimist's misread: "some spacing bugs in prototypes, who cares." Also wrong. The violations weren't exotic edge cases. Typography, spacing, unapproved color: the basics, the things the format existed to prevent. If the floor can crack there, it cracks everywhere.

The honest reading: structured context raises the floor. It does not hold the ceiling. Volume is a multiplier on both output and drift, and at 1,000x volume, a 1 percent violation rate is a violations pipeline. The interesting question stops being "can agents generate on-system UI?" (mostly, yes) and becomes "what catches the drift, at machine speed, without turning humans into full-time inspectors?"

The answer has a shape, and it's a loop: observe, detect, suggest, fix, with a human in the oversight seat. Observe what agents actually produce, continuously, not in quarterly audits. Detect violations mechanically: if your tokens are data and your component contracts are schemas, then "does this output use approved tokens" is a lintable question, which means it's a CI question. Suggest the correction, because most drift is mappable back to the nearest legal value. Fix automatically where the fix is unambiguous; queue for a human where it isn't. The human stops reviewing everything and starts reviewing exceptions, which is the only version of review that survives agent-scale volume.

Notice what that loop requires: the same machine-readable substrate that enabled the generation. You cannot lint against a style guide written in prose. Tokens-as-data, contracts-as-schemas, specs-as-files aren't just generation inputs. They're the assertions your governance runs on. This is why "add AI generation" and "add automated governance" aren't two projects. They're one project with two exits, and teams that build the first without the second are choosing the 4,300 problem on purpose.

There's a cultural half to this too, and it's the part I'd underline for design system leads. Your job description quietly changed. The old failure mode was neglect: the system drifted because nobody used it. The new failure mode is unsupervised success: the system gets used ten thousand times a month by consumers that never sleep, never push back, and never ask "is this right?" in Slack. A design system team in an agentic org is a governance team, whether the org chart says so or not. Budget for it. The people doing this seriously put ongoing maintenance at something like a third of the effort, and anyone selling it as free is not running it.

So: generate boldly, and instrument everything. Celebrate the 4,300. Audit the sample. Wire the loop. The teams that win this era aren't the ones whose agents make no mistakes. They're the ones who catch drift in CI instead of in production, and whose humans spend their judgment where judgment is actually needed.

The free starter kit at tokenstoagents.ai/kit includes the governance-loop checklist: what to observe, what to lint, and what still needs human eyes.