12 agents handed over, no licence fees ever

A data centre brand site, plus the AI engineers that made it

3 daysper website build, down from about three weeks
12agents built, handed over and still in use
8 of 16touch points now fully agentic
0licence fees, ever
Client
Creative agency, 30 people
Delivered
August 2026
Read time
2 min
Reference
On a call, under agreement
StackTypeScriptSQLReactNext.js 16Payload CMSNeon PostgresTailwind CSSRadix UIGSAPMotionVercelGitHub

Who

A thirty-person creative agency in Australia, building a website for a Melbourne data centre brand. I mapped how they build client sites, rebuilt that workflow as a team of twelve agents, and trained their engineers to run it.

The problem

Every client website passed through six people and sixteen touch points. Turning a signed-off Figma design into a live site took about three weeks. They wanted this one fast, and they wanted their own team to build the next one. Most agencies hand over a website. Nobody hands over how it was made.

What I built

Stage What I did Output
Map Talked to the people who do the work, and mapped the build end to end. 6 roles, 16 touch points
Gate Decided for each touch point whether it should be an agent, an automation, or a person. 8 agentic, the rest automated or human gated
Build Built a team of twelve agents as a Claude Code plugin. It reads designs straight from Figma and ships a Next.js site to Vercel. Figma design to live site
CMS Set up Payload, an open source CMS, to fit the brand and the CMS skill their own clients have. A CMS their clients can run
Handover Installed the team in the agency's own GitHub organisation with full IP. Trained them with documents, live sessions and video. Their engineers run it

Results

  • Three weeks became about three days for each website build.
  • Eight of sixteen touch points run fully agentic. The operator's job is now mostly checking.
  • The agency owns the IP and their team still runs it without me.

What they own now

All of it. The agents sit in their GitHub, in their name, and we kept nothing they need. Their engineers picked it up in one session, were suggesting their own improvements the same week, and now build with it on other work.

The first step

Name the task. I’ll tell you if AI should touch it.

Eugene Zhang, founder of Bon LabsFind your first AI build