Four meshed gears of different sizes drive a small printing press, the smallest spinning fast at the centre and the largest turning slowly at the edge, while a figure seen from behind reads a pressure gauge on the frame, the whole scene drawn with wide empty margins on every side.

Design Systems Collective

Four feedback loops for an AI-native design system

Melina Kukkasela · 25 October 2026

How-to · design systems · agents · Agent context files

Takeaway

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.

Summary

Melina Kukkasela, Lead Product Designer at the software company Visma, develops Adesy, a design system shared by several brands and front-end frameworks. Her team built four feedback loops, from checks on every run to monthly scored tests, to learn whether changes to its AI agent instructions and documents helped. The first scored round found that agents used components far more accurately with the setup, though no run passed every check.

Key points

Seven AI agents help build Adesy, covering research, design, engineering, accessibility, Figma, testing and a lead that routes the work.
A repeated lesson becomes a one-line rule, which the testing agent checks and a person approves in the pull request. After one run's lessons became rules, the next full run used about 78% fewer tokens, on a different component.
In 18 runs with Claude Code, agents with the Adesy setup used components correctly in 67% of runs, against 22% without it.
The failed runs showed that the documentation server wrongly said 14 child components did not exist. Nine of the team's own checks also raised false alarms.

Implication

A design system team can measure whether its AI agent instructions help by running real tasks with and without them, then reading the failed runs. Kukkasela states that this is one practitioner's case with a small sample, so the figures show a direction rather than proof.

Plays

Engineering

Give a coding AI agent the same build tasks several times with the design system's instructions and documents and several times without them, then score every result with pass-or-fail checks.

Design

When the same lesson recurs across an AI agent's runs, write it as a one-line rule in the agents' instructions, and have a designer approve the rule in the pull request.

Derived by Working Surface from the article; more in the Playbook. Source line: The score told us that something failed, and reading the failed runs told us why.

Source issue

26 October 2026: teams now judge AI work by tests a person sets, from pass-or-fail checks on what agents build to plain questions about who asked for a feature.