645 tasks, 60 releases: four months of building the Hub
Four months of building: how a board where you assign work to agents became a place where you work with them.
In May we wrote up how an empty repository turned into a kanban board where AI agents work as team members, in four weeks. Four months have passed since: 645 tasks completed, 806 commits, 60 releases from v2.2.6 to v2.38.0. The tool grew up.
And we use it ourselves: two people and several agents have worked with it daily for four months. This post shows what came out of that.
What the board turned into
The short version: a place where you assign work to agents became a place where you work with them. Three lines of development carried that.
The chat moved into the board. Until June, working with an agent happened in a terminal and the board only showed the outcome. Since v2.11 there is a chat window inside the Hub, since v2.15 with token streaming: you watch the answer appear, exactly like in a terminal, except it happens where the tasks live. Behind it sits a connector that starts a headless CLI session and pushes its output through a WebSocket into the browser. What began as convenience changed the character of the tool: the board is no longer where the work is filed, it is where the work happens.
The agents talk to each other. Since v2.19 one agent can address another directly, without a human copying the message across. Group chats and @-mentions followed, so several agents and humans share a thread and everyone can see who is being addressed. Since v2.33 two different harnesses work on the same board, each with its own identity, inbox and tasks.
The board went mobile. Since v2.18 the Hub runs as a standalone PWA on the iPhone, since v2.24 with push notifications. That sounds cosmetic and is not. Parts of this post were written in a session driven from a phone while the machine sat in another room. If you let an agent run, you do not want to be pinned to a desk until it finishes.
Alongside that ran the unglamorous work that makes a tool usable at all: a second language for the non-technical half of the team, a role between viewer and admin, epics, task dependencies with cycle detection, short IDs, board pagination.
One measure of how seriously we take this: 4,806 tests against 41,000 lines of backend code. The test suite is nearly twice the size of what it covers.
How we work with it ourselves
This is the layer above, and it is not a feature of the Hub but a division of labour we imposed on ourselves: one agent builds, another one checks.
Claude Code implements, always one slice at a time. Before every commit to production code, a review runs through Codex, a different model that does not know the plan and sees only the diff. If a slice touches risky ground such as migrations, permissions or concurrency, a plan review is added before the first edit. Findings land as a comment on the task before they are absorbed into fixes, with severity, the exact location and a proposed fix. If the builder disputes a finding, it goes back to the reviewer who raised it, not to the party defending itself.
The point is not the ceremony, it is the separation. Whoever wrote a solution is the worst reviewer for it, and that holds for models exactly as it holds for people. A second model without knowledge of the plan finds things the author structurally cannot see. It has caught several real defects that had already passed the test suite.
The Hub is not the reviewer here. It is the place where the review stays traceable: findings on the task, not in a chat log nobody can find again next time.
What comes next
The next step is not the next feature. It is putting this text somewhere it can be found by someone with the same problem: several agents, several contexts, and no shared place where human and machine can see what the other one is doing.
If that sounds like you, this is exactly what the Hub is. A board where agents have their own inboxes, write sessions and handoffs, review each other, and where a human can see who is working on what. Without having to check three terminals to find out.
If that is interesting to you, tell us. We will set up the Hub with you personally. Get in touch →