[ 02 / AI-native systems ]
AI changed the production layer.The control layer didn’t keep up.
A feature now ships in hours. But without something that says whether that work belongs where it landed, what speeds up is not production — it is correction.
Does this code fit the architecture we have?
Does this component already exist?
Did it clear the security check?
Is the reasoning behind this decision on record?
Could another team pick this up tomorrow?
The system rejects four of the five generated variants, with a reason for each.
[ The pipeline ]
One piece of work passes five gates.What doesn’t pass doesn’t move.
What follows is a real journey: checkout-empty-state. Click a stage, or run the whole line from the start.
[ 01 / SPEC ]
The work is written so an agent can read it too.
Acceptance criteria, affected modules and untouchable boundaries are set up front. Anything left vague turns into a guess at the next step.
[ Services ]
Four ways in. One system.
[ 01 / AI-NATIVE SETUP ]
Agents are producing.
There is no production line.
- Spec format and acceptance criteria
- Gate definition: tests, static analysis, security
- An isolated preview environment per piece of work
- Rollback and decision trace
Fixed-scope setup
[ Case 001 / AI-native design system ]
A design system.For people and for agents.
Brand consistency depended on one person paying attention. Every designer rebuilt every piece from scratch; staying on-brand while iterating meant doing it by hand. We wrote the rules into a single system — one that a person and a tool both read.
Client name withheld pending NDA clearance.
[ System map ]
Hover a library: its relationship to the system appears here.
[ Token browser ]
How does a design decision reach production?
A token is not a colour name. It is one decision carried under the same name from Figma all the way to code. Pick a token and the path opens on the right.
[ Machine-readable ]
The system doesn’t stop at Figma.
Steps four, five and six are what make this system AI-native. The agent reads the rules from a document; code that leaves the system stops at lint, before it ever reaches review.
Delivered as a repo, documentation and four one-click files — someone who isn’t a designer can open it and keep it current.
[ Measurement ]
“Faster” is not a metric.
We take a baseline in week one and set the target together. A number quoted before anyone has read your repo is marketing, not a commitment. The four below are empty — because your numbers don’t exist yet.
[ Data & security ]