Auto Kanban for AI Coding Agents: Tasks That Start on Their Own

What the Auto lane is

The Auto lane is a column in the Kanban, between Pending and In Progress, marked with a lightning icon. Tasks in it are waiting to be picked up. As soon as there is a free slot, the app opens a session for the next task and moves it to In Progress.

Everything else about the board works as before: projects, labels, subtasks, and agents that update their own cards. If you are new to the board itself, read the task management guide first. This one only covers what the Auto lane adds.

How to queue a task

There are three ways to put a task in Auto:

• Drag a card from Pending into the Auto lane. It uses your Auto defaults.
• Click the lightning button on a Pending card ("Run automatically"). It opens the settings for that task only. Change what you need and click "Queue task". Your Auto defaults stay the same.
• When you create a task, turn on "Start automatically". The task is saved straight into the Auto lane and starts as soon as a slot is free.

The task needs a project, because that is where the session runs. If it has none, the app asks you to assign one before it can go into Auto. You can also queue tasks from the mobile app, and an agent can create a task that is already queued when you ask it to.

The settings each task carries

The "Run automatically" settings travel with the task, so every card can run differently:

Work style is how much the agent should stop and ask. "Normal" adds nothing to the task. The other two add a short instruction, and you can read the exact text under "Instructions added to the task" before queueing. "Use best judgment" is the one for tasks you want finished without questions, while "Ask only for blockers" still stops when a missing fact or decision really blocks the work.

Worktrees are used only when the project supports them. If it does not, for example a repository with no commits yet, Auto uses the project folder and tells you why. Why worktrees matter when several agents work on one repo is explained in the Git worktrees guide.

How many tasks run at once

By default there is no Auto limit: every queued task gets its own session when there is room, in the same project or across projects. If you want fewer at a time, open Auto settings › Advanced settings and set "Maximum tasks in parallel" to Custom with a number. That maximum applies to all your local projects together. Lowering it never interrupts work that is already running.

Your plan's session limit still applies. The lane header shows the live count, for example "2 running · 3 queued", plus the maximum when you set one.

Pickup runs every few seconds, even with the Kanban window closed, so you can queue work and forget about the board. If you use a daily budget with "Pause new work", tasks for a provider at its limit stay queued and the header tells you they are held by the usage cap. They start again when the new day opens.

What the task session receives

Each task opens a Chat session with the agent and settings on the card. The first message is the task itself:

• The task number and title.
• The description, if the task has one.
• Any images attached to the task.
• The work style instructions, if you chose "Ask only for blockers" or "Use best judgment".
• A final line that marks the task as In Progress and links it to that session.

That is all. The more useful the description, the better the result, so write tasks as if you were handing them to a colleague: what to change, where, and how you will know it works.

While it runs, the card in In Progress has a "Go to session" button that takes you straight to the conversation, so you can watch or answer a question.

When a task finishes

The agent follows the normal task flow: when it is done it writes its summary and moves the card to In Testing. Auto never marks anything Completed. That step stays yours, after you review the result.

The session stays open after the task, so you can read what happened, ask for a change, or commit. The slot is freed and the next queued task starts.

When things go wrong, Auto is careful not to run the same task twice:

• If you close a session before the task is finished, or the session fails to start, the card stays where it is and leaves Auto. It does not retry on its own.
• If the app restarts in the middle of a task, it does not relaunch it, because that would open a duplicate conversation. You can queue it again or resume the conversation yourself.
• A session waiting for your answer keeps its slot while it waits.

A practical workflow: queue the backlog, review in the morning

The Auto lane shines with a list of small, well described tasks. A routine that works:

• At the end of the day, write or clean up five to ten small tasks: bug fixes, tests, copy changes, small refactors.
• Pick the agent per task. Claude for the tricky refactor, Codex for the tests, a faster model for copy changes.
• Use a Git worktree for each one so they never step on each other.
• Set the work style to "Use best judgment" for tasks you want finished without questions.
• Queue them, and set a daily budget so the queue cannot eat tomorrow's quota.
• In the morning, open In Testing and review each card with its summary and its session.

When the work is one big goal instead of a list, with parts that depend on each other, a coordinator is the better tool. And to compare this with other ways of running agents side by side, see the AI CLI agent swarm guide.