Design system teams now judge their work for AI agents by testing what the agents actually build, and they keep the final decisions with named people. A design lead at a financial technology company adds that an AI feature still needs a person with a problem before anyone builds it.
Melina Kukkasela (26 Oct · 01) is Lead Product Designer at the software company Visma, where she develops Adesy, a design system shared by several brands and front-end frameworks. To learn whether changes to Adesy's AI agent instructions helped, her team gave a coding AI agent the same three tasks with and without the Adesy setup. With the setup, the agents used components correctly in 67% of runs, against 22% without it, yet no run passed every check. Reading the failed runs showed why: the agents skipped the grid documentation, and nine of the team's own checks raised false alarms. Her point is that the score shows something failed, while only the failed runs show what to fix.
Cassie Groos (26 Oct · 02), a design system practitioner, reported on Converge 2026, a design system conference held in Newcastle by the documentation company Zeroheight. Dan Donald, one of the speakers, sorted guidance into three kinds: rules that agents follow and that are checked, guidelines that people review, and decisions that people own by name. Zeroheight's founder, Jerome de Lafargue, described an agent that proposes documentation updates when a Figma design changes, with a person approving each one. Groos drew the common thread herself: "However much the agents do, somebody still has to decide."
Jon Sabutis (26 Oct · 03), who leads design at a financial technology company, argues that many products ship the same AI chat panel. The panel is switched on by default and hard to turn off, because companies bought a capability before finding a problem for it. Before any AI item is built, he asks who asked for it and what it is named after. He also asks whether it would ship switched off and whether it would pay at full cost. He then describes how a designer without budget authority can bring those questions to a roadmap meeting, for example by offering to map three real tasks first.
All three pieces put a test in front of AI work: pass-or-fail checks for what agents build, and plain product questions for what a team ships. In each case a named person decides what counts as good enough.
Kukkasela's figures rest on 18 runs of three tasks, and she says they show a direction, not proof. The conference speakers' models are reported second-hand, so the strongest claims have not yet been tested beyond the teams that made them.<br />
Melina Kukkasela at Visma tested coding AI agents with and without her design system's setup, and the failed runs showed which documents and checks to fix.
At Converge 2026, design system speakers proposed firm automatic checks for AI agents and named human owners for decisions, according to attendee Cassie Groos.
Jon Sabutis, a design lead at a financial technology company, argues that AI features built without a user problem drive people away, and offers four questions to ask first.