The Designosaurs™
As of 2026-09-23, the Designosaurs are behind-the-scenes agents in our factory architecture — clear, fun, and deadly serious ways to segment responsibilities the way a human software organization assigns them: security, privacy, ethics, efficiency, deployment, whether to build at all, and what it should look like. The names are how we remember who owns what.
How far to show them is an open question. We envision two modes of operation — functionally the same, dramatically different in personality: a dry, efficient mode where the questions and the workflow stay plain, and a memorable mode where Cleo and Dromeo speak in the interface. Neither mode is built; we will let partners choose.
On record is an idea we have not scoped: a parent-and-child experience that teaches how agentic AI works, with the Designosaurs as guides — Designosaur Academy, an idea since February 2026.
Four of the nine roles are partial today; the cards say which. No video autoplays and nothing here tracks you.
PARTIAL
Cleo · Strategy and intake
Are we building the right thing?
Take in what a customer needs by conversation before anything is built.
Question on intake: are we building the right thing?
The intake wizard authors a requirements document by conversation and shows it back.
Partner notes reach the same queue as my own requests.
The full line has not run clean end to end.
PARTIAL
Dromeo · Design
Give me a head start on how we want this to look.
Turn requirements into a declared design in three looks.
Every shipped app carries a Trio of runtime-switchable looks.
The theme engine runs with a library of 25 looks.
A studio’s public site can be read into a partial design.
The design stage is wired into the line; end-to-end generation is not yet proven.
PARTIAL
Val · Ethics and scope
Ethics, scope, and whether we should build anything at all.
Check scope and ethics before work proceeds.
Question on scope: is something already out there we should use instead of building?
The ethics gate exists in the line; the line has not run clean end to end.
The gateway’s policy admits models from democratic countries only.
PARTIAL
Ada · Code
Code that survives real checks.
Write code against real checks and repair it from the checker’s own diagnostics.
The code has targets for web, iPhone and Android.
Ada builds one screen per run today.
There has not been a clean end-to-end run yet.
LIVE
Trully · Acceptance and design review
Is this good enough to ship?
Review the result, score it and ask a person when unsure.
Machine checks run first; a person judges what they cannot decide.
Acceptance belongs on the test tier; review of generated screens remains partial.
LIVE
Tycho · Records
What was decided, checked and promoted.
Keep the record of what was decided, checked and promoted.
The decisions ledger and promotion records keep decisions and releases traceable.
The wiki is compiled from sources.
LIVE
Sun-sin · Ladder and deploy
Deployment only when checks and promotion records say so.
Move work from test to production only by a recorded promotion.
The environment ladder requires checks and a promotion record to reach production.
LIVE
Penny · Health and spend
Efficiency: health, spend, and oracles before judges.
Watch health and spend, and say so before it hurts.
Deterministic oracles decide what they can before any judge is asked.
Spend policy and the health board make costs and service health visible.
The weekly goal review compares progress with the recorded goals.
LIVE
Wolfgang · Conductor
Who owns the line and the schedule.
Conduct the line and own decisions and scheduling.
The stage dispatcher moves work from one role to the next.
The fleet merges packs of work under gates; the generation line remains partial.
Security: production state is commanded through grant-locked doors; raw writes to the six work-order carriers are refused at runtime.
Privacy: model origin, weights license and endpoint country are routing attributes; personally identifiable information does not leave allied endpoints.