Acceptance: checking an agent’s output against your design system

A check written after the work is a matter of taste. A check written before it is a standard. For each kind of work your agents do, list what must be true before it ships, and tie each line to a rule the system already states: this design token, that component, this line of the guidance. A line with no rule behind it is a sign the rule is missing. Write the rule first, in your guidance for agents.

Anything a script can decide, a script should decide, because it decides the same way every time:

• Values written by hand. A checker lists every color, size or font typed where a design token belongs. The Simbolik Design System’s checker does this, and also lists components a site rebuilt by hand. For any CSS, Stylelint’s color-no-hex rule flags hex colors.
• Props that do not exist. With typed components, a type check fails on a variant the system does not have.
• Accessibility. axe-core tests rules from WCAG 2.0, 2.1 and 2.2, and marks what it cannot decide as incomplete for a person.
• Visual change. A screenshot compared against a saved reference, as Playwright does, shows what moved since the last accepted version.

Between the scripts and the person, a second agent can review the work against the written rules and send back what is off. Simbolik’s service page describes this layer as review agents that find what is off and fix it, and its AMPECO case describes an auditor agent that checks every output before designers see it. It catches what a script cannot express and a person should not have to.

The person who signs off should meet only work that already passed. Their job is what no tool decides: whether the hierarchy is clear, whether the words fit, whether the screen looks like your product and not like anyone’s. Tools cover part of accessibility too: axe-core’s makers say it finds about 57% of WCAG issues on average, so keyboard use, focus order and the meaning of labels still need a look against WCAG 2.2.

A failed check is an instruction. Give the agent the finding itself, ideally the checker’s JSON report, and let it fix and run the checks again. When the same failure comes back twice, the problem is upstream: a missing design token, a rule that is unclear, a component that is hard to find. Fix that, then add a prompt for it to your test set of prompts for AI agents, so the fix stays fixed.

Acceptance is also where a custom AI-ready system earns its keep: its checks know your brand, not a public kit’s. Work with Simbolik (opens in a new tab) when you want them built, or see which public systems give an agent the most to check against on the best AI-ready design systems, ranked.