An agent team with a kanban board and a reviewer

I run my business alone, and the work spreads across planning, copy, design, code and legal questions. Done one after another, that takes weeks, because every switch means getting back into context. I wanted to find out whether a team of AI agents with clear roles can split the work the way a small office does: one plans, one designs, one builds, one checks. I did not test this as an exercise. I ran it on the first day of the board, with real tasks from my business.

Seven roles work through one shared kanban board: business, design, product planning, build, copy, review, and a coordinating role that creates and assigns cards. Every card has a brief, dependencies and a result. A card only starts once the cards before it are done. The check is always done by a different agent than the one who built the thing. On that first day there were two independent reviews and three rounds of corrections. Hermes Agent and Claude Code run the roles, and the board lives in a local database.

• Hermes Agent
• Claude Code
• Kanban board
• SQLite
• Python

On 12 October 2026, the board's first day, I created 30 tickets. 29 of them were closed the same day. The 30th was parked on purpose until I made a decision. Seven roles took part, and the tasks ranged from an explainer video and a test app to a legal check. The number comes from an export of the board database, and I counted it twice independently. How many hours this saves compared with doing it by hand, I have not measured, so I make no claim about it.

A video of about a minute shows the demo board Atelier Beispiel, an invented photo studio with nine sample tickets. You see cards move through the columns, a reviewer send a card back, and the correction pass afterwards. Everything in it is sample data, not a real job. The role list with avatars comes with it. The count behind the number contains titles only, no names, and I can show it to you on request.

This is not autonomous. The coordinating role kept steering throughout the day, one card stayed blocked, and two parallel design sessions created a duplicate that I had to close. Decisions about direction, sign-off and money stay with me. Whether the results are better than what I would have managed alone, I have not measured. The screens and video in this case come from a demo board with sample data, not from my real board, because the real one carries client and project names.