Eleven agent roles, one human, and a queue worked in public.

The product is built by AI agent sessions playing defined roles under a written rulebook, with a human (Dinis Cruz) making the decisions only a human can. Every role's outputs are committed; the queue below is live.

open
blocked
done
The principle

Roles, not vibes

An agentic team without structure produces the same mess as a human team without structure — faster. The structure here is simple: each role has a written ROLE file (mandate, boundaries, outputs), a folder for its reviews, and the rulebook binds them all to the reality discipline. When a role has produced nothing, its empty folder says so — the structure is honest about its own exercise.

RoleOwns
ConductorRouting every task across roles; priorities
ArchitectThe load-bearing decisions: no backend, key-as-billing, privacy modes
DevThe product front end, built from the existing SGraph tools
DesignerThe one-motion experience — form, function, intention
DevOpsThe two-branch CI, auto-tagging, both deploy estates
LibrarianReality vs proposed; the library; the master index
QATest strategy; the audio-format matrix is test surface #1
AppSecKey handling: caps, rotation, revocation, the seeded-key abuse model
DPOThe privacy-mode claims — routed / restricted / browser-local must be true
AmbassadorGrowth and positioning: sell the availability gap, honestly
JournalistThe site's Updates section; the per-version changelog, with the Librarian
Live

What the team is working on

Open and blocked issues as of this deploy. Each is a markdown file in the repo — click through for the full brief, status history and outcome criteria.

#TitleStatePriorityCreated
loading issues.json…

The full queue including everything done is on the hub page.

Receipts
  • team/README.md — the roster and the three-team model
  • The rulebook every agent session starts from
  • issues/ — the queue itself, files in folders, no tracker service