We accidentally spent six years preparing for AI.
The honest version: what worked, what broke, and how a fashion brand with no engineering team ended up running on a company brain. No vanity metrics, no "10x productivity."
We documented everything. We just didn't know why yet.
We're an asynchronous-first company. The founder hates meetings, loves working weekends, takes Mondays off. To make that culture work, everything had to be written down — emails instead of meetings, Looms instead of briefings, a clean Drive, a clean Notion, manuals for everything from refunds to photoshoots.
That was a productivity decision, not an AI strategy. But it quietly became the most valuable asset we own.
One founder, one WhatsApp agent, one obsession.
When agent platforms got good, the founder built himself one over a weekend — a personal agent living in WhatsApp. It was mind-blowing for exactly one person: him. The obvious next thought: "I need this, but for my company."
First target: customer service. We gave an agent an approved corpus from two support inboxes, plus the operating manuals and website, and used that material to build a governed knowledge base and decision tree.
It built its own knowledge base and decision tree. Honest admission: we couldn't tell you exactly what it built. But it works — it answers customers autonomously, and once a week it re-reads the new conversations and updates its own knowledge.
Nobody cared — until the agents joined Slack.
For weeks the founder emailed the team about AI. Dashboards, demos, links. Nobody answered. We work in fashion; the team is creative, not technical, and skeptical of AI for good reasons.
What finally worked wasn't a memo. We put the agents inside Slack, with human names and personalities, and the founder simply started talking to them in the group channels like colleagues. People asked "who's Strategy?" — then started DMing the agents themselves. Nobody was told to.
Then we made it stick with mechanics, not speeches: part of the bonus is tied to AI usage, there’s a monthly award for the best AI solution built by an employee, and since May the AI client is mandatory — like email, like spreadsheets. Culture change needs incentives, not memos.
Then we scaled it — and everything broke.
The team's requests got more ambitious: fill this sheet, draft this email, fix this report. Three agents sharing all the context started mixing it up. The CS agent once emailed a customer about an internal spreadsheet. The founder became the IT department, fixing agents instead of running the company.
Two fixes came out of that period. First: one agent per domain — customer service, finance, retail, merchandising, marketing, people — each owning its lane, each with its own context. Second: when AI desktop clients became usable for non-technical people, every department got one. Today the team opens it first thing, like email.
The honest part: something still breaks most days. It usually takes ~20 minutes and a voice note to fix. We publish that too.
One brain. Every agent and every employee, reading and writing.
What emerged is the architecture we now publish: a central brain — structured, versioned markdown, indexed and shared — with every agent and every employee's AI client connected to it as nodes. Everyone reads the same memory. Everyone writes back what they learn.
Meetings are recorded, transcribed, and written to the brain. Inboxes get swept daily and summarized. Bank accounts reconcile against invoices every six hours. The world model refreshes weekly from real signals. When the founder says one thing on Monday and another on Friday, the team searches the brain and quotes him back.
That's the unlock we'd point any operator to: formalize the context, and execution follows. A video call became a predictive P&L. A finance lead with zero code background shipped her own treasury dashboard. None of that is a model feature — it's what a model can do when it actually knows your company.
Running in production. Still imperfect. Fully public.
Today: seven production agent runtimes plus a founder command center, models swappable by design, 97 authenticated MCP tools, 4,842 Brain documents, 373 available skills, 47 canonical skills and more than 42,000 action receipts for ~€631/month all-in. Our July due-diligence scored Brain maturity at 6.8/10 and broad autonomous execution at 2/10: the bottleneck has moved from capturing context to verified closure. The strengths, gaps, architecture contract and closure-first pilot are all documented in the playbook.
We published the whole thing because we wish it had existed when we started. Read it, fork it, adapt it. If you ship something, tell us.