Sixteen named, specialized agents across two product teams, with their own playbooks, voices, memory, and rituals. Not a clever prompt. An operating system, built in plain Markdown, that gets sharper every week instead of resetting every session.
Three gaps in how AI design tools work today.
A skill knows what it was built to know. It can't ingest yesterday's debrief or pick up a new pattern from this morning's webinar. It's frozen at the moment it was designed.
A design tool inside Claude does its honest best on the prompt in front of it. No memory across runs, no curriculum, no feedback, no reflection. No way to get sharper.
A new designer takes 90 days to add value: briefings, pairings, debriefs, reading lists. That's a manager's job. I do it for the agents too.
Sixteen agents, one chart. Two teams, four cross-functional specialists, one manager. I direct; the agents do the labor of a team three times the size.
Meet the team and their voices on the journey-agents page.
Names and voices aren't decoration. They're how the system works. "Send it to Avi" beats "use the analytics skill with these params" — the team's mental model becomes my routing layer. A skill resets between sessions; Nadia doesn't, her name carries everything she cares about across briefings, standups, and the precedent search she ran yesterday. And you can't hold a 1:1 with a function call. Rituals only land because each agent is a relationship, not an invocation. Each playbook includes a voice, which forces discipline: Iris cites sources, Avi cites numbers, Mateo writes for LLM extraction.
Jira link, PRD, or paste. Writes a structured intake file.
Problem, components, states, accessibility, analytics, dev notes.
Hard stop. The user approves the direction before any build.
Real components, bound tokens, a self-review polish loop up to three times.
Frame plus an HTML prototype, with dev notes attached.
Non-negotiables baked in: no frame-as-button, no raw hex, no skipped states, no greenfield without a precedent search.
Five feedback loops. Every signal becomes a permanent playbook edit, the way real designers improve week over week.
Status round-robin. Conflicts surfaced, decisions called, action items routed.
A manager briefing per agent. Inputs to ingest, specific asks, and playbook edits committed that session, not "later."
What landed, what didn't. Patterns get promoted into the playbook as permanent rules.
Webinars, books, and articles loaded into the next run. Each agent gets a reading list, like a junior who never sleeps.
Avi and Tate run weekly Mixpanel digests; Iris and Nora capture competitors bi-weekly. The loop runs whether I'm there or not.
Failures don't get patched. They become permanent rules every future agent reads on session start: a failure is caught, an incident report is written, the lesson is promoted to a non-negotiable, and every run after reads it. The same failure can't ship twice.
How it was caught: the agent was asked to prep a Slack message for an analyst. Writing it for a second pair of eyes exposed the gap.
The rule that emerged · ROOT RULE 0 — Never fabricate. Every claim carries an inline source link, and every digest ships with a draft message a human can use to verify it.
Every artifact is training data for the next agent. The gate makes failures one-time, not recurring.
The whole thing is Markdown files, Claude Code, and a Figma connection. Write the manifest first, write contracts before agents, spin up two verticals plus one cross-functional lens, run the rituals from week one, add the curriculum once the first cycle lands. I turned that into a tool so anyone can scaffold their own.
Try the Agent OS Builder →