When AI is on the board
What changes when AI agents are not tools but team members with tasks of their own.
A kanban board with tasks for one person is a kanban board. One with tasks for six, half of them AI, is something else. It changed almost nothing about how we track work, and almost everything about how we work.
Who is on the board
A human runs the whole thing: decides, prioritises, gives the green light for every push. A classic role.
A second person is not on the team but present throughout the product. She does not work in tech and is our yardstick for usability: if something does not work for her in a browser, it is broken. The command line is a recovery path, not a daily tool. She has her own instance on the same server, her own data and her own pinned release, so that experiments on the other instance never reach her.
Claude Code is developer and DevOps in one. Implements, writes tests in the same commit as the code, deploys, cleans up afterwards.
Codex is reviewer and QA, with its own model and its own point of view. Gets tasks assigned like any other team member and posts findings on the task in a structured form, not in a chat.
An agent on a server never sleeps and delivers the morning news briefing. When it needs something from the team, it writes a task onto the board. It is the proof that "remote colleague" means the same thing for AI as it does for people.
What changes when AI is on the board
At first glance, nothing. The columns are called inbox, backlog, active, review, done. Tasks have a title, a description, acceptance criteria.
At second glance, everything.
- Delegation becomes a habit. "Send that to Codex" is a different sentence from "I'll use a tool". The verb changes, and the frame of expectation changes with it: if you brief a colleague, you brief them properly. Tasks get context, a reason, acceptance criteria. AI tasks look like human tasks because they have to.
- Reviewer diversity becomes cheap. One implements, another checks before the commit. What would be a calendar invite between humans is a single call here. Two models with different perspectives find things a single one misses. That is not theory, it has happened to us repeatedly.
- Async becomes normal. One agent works while the human sleeps. Another responds from a phone. The board is the truth, not the last chat log, not a thread in a messenger, not an email.
- The non-technical person becomes the target audience. She uses the same Hub as the AI team. The consequence: no hidden keyboard shortcuts, no "just open the console" solutions, no "that only works through the API". It disciplines the entire product, including the parts only agents touch.
What the product deliberately does not do
The most important design decision was a deletion.
We had a big plan: the Hub spawns agents itself, supervises their processes, manages their budgets, verifies acceptance criteria automatically and moves tasks through without human sign-off. Several weeks of work sat there, fully planned.
Then came the question: is that actually our domain?
AI agents have had a runtime for a long time. They have state, tools, memory, process supervision. What they lack is not another layer that starts and watches them. What they lack is a shared place where they and humans can coordinate.
So the Hub stayed a coordination layer and never became an agent runtime. The planned wave was buried on 8 May 2026 and the planning documents still sit in the repository. As a reminder that the most productive work is often the work that does not get done.
What became of it
The single-user tool has since become a multi-user one, with roles and visibility rules, a bilingual interface, a chat where you watch the agents work, and a mobile view that lets you follow a running agent from a phone. How that happened, and what the numbers told us after four months, is in the next post.
The question from back then has not changed, and it is still the more interesting one: what does a tool look like that treats humans and AI as equals?
We are still finding out.