Task boards

Claude Code kanban: a real task board for AI agents, column by column

Updated November 17, 20268 min read

A Claude Code kanban board is a task board where each card is one job for an AI coding agent, and the columns show how far each job has got. Instead of explaining the idea in the abstract, this page walks through one example board from left to right, shows how an agent picks up a card, gives you a TASKS.md file you can copy, and then compares the ways to run a board for coding agents.

The example: a booking site for a hair salon

Say you are building a small booking website for a hair salon. Customers pick a service, choose a time and get a confirmation email. You have two Claude Code agents and want to keep both busy without losing track of what each one is doing. A Claude Code task board is the simplest way to manage tasks for coding agents like this: each job is written down once, and its column tells you who has it.

The board has five columns: Ideas, Ready, Agent working, Review and Done. The extra Review column matters. AI agents finish fast and sound confident, so "the agent says it's done" and "it is actually done" need to be two different places on the board.

Walking the salon board from Ideas to Done

  1. 1

    Ideas: anything you might want

    Rough notes go here, with no detail needed. "Text reminders the day before." "Let staff block out holidays." "Gift cards?" Nothing in Ideas is ready for an agent. It is a parking spot, so good thoughts do not get lost while you work on something else.

  2. 2

    Ready: cards an agent could finish alone

    A card moves to Ready only when it has three things: one clear goal, where in the project the work happens, and a short "done when" list you can check. "Show available times for a chosen day" with "done when: booked times are greyed out" is Ready. "Make booking better" is not. Order this column with the most important card at the top.

  3. 3

    Agent working: one agent per card

    When an agent takes a card, it moves here with the agent's name on it. Keep it to one card per agent. Two cards that edit the same files should not sit in this column at the same time, or the agents will overwrite each other. If both must run, give each agent its own copy of the project with a git worktree.

  4. 4

    Review: the agent stopped, now you check

    The agent moves its card here when it thinks it has finished. You go down the "done when" list and try each line in the app. If something fails, write a short note on the card and move it back to Ready. That note becomes part of the next prompt.

  5. 5

    Done: checked, merged, finished

    Only you move cards here, after the work is merged into the main project. If the agent mentioned anything new along the way, like "the email template has no plain text version", write it as a fresh card in Ideas before you forget.

How an agent picks up a card

Claude Code does not watch a board by itself. Something has to hand it the card. There are three common ways:

  1. You paste it. Copy the card into the agent's prompt box. Simple, and it works with any board.
  2. The agent reads a file. If the board lives in a file inside the project (see the TASKS.md example below), you can tell the agent to read it and take the top card.
  3. A board tool starts the agent. Some boards launch an agent straight from a card, so you never copy anything.

For the file approach, a prompt like this works well:

Read TASKS.md. Take the top card under "Ready", move it to
"Agent working" with your name, and do it. Follow the "Done when"
list. When you think it's finished, move the card to "Review" and
write one line about what you changed. Don't touch other cards.

Claude Code also keeps its own task list inside a session. With the experimental agent teams feature, that list is shared: tasks are pending, in progress or completed, can wait on other tasks, and teammates claim free tasks by themselves. Ctrl+T shows it. It is a good fit for splitting one job, though it lives in one terminal session and covers Claude only. See Claude Code subagents for the simpler helper version.

A TASKS.md board you can copy

The cheapest kanban board for AI agents is a plain text file in your project. Agents can read and edit it like any other file, and you can see every change in git. Here is the salon board as a file:

# TASKS

## Ideas
- Text reminders the day before
- Staff can block out holidays
- Gift cards

## Ready
### Show free times for a chosen day
Where: booking page only. Don't touch payments.
Done when:
- Picking a day shows 30-minute slots from opening to closing
- Booked slots are greyed out and can't be clicked

### Confirmation email
Where: the booking form and the email code.
Done when:
- Customer gets an email with service, date and time
- A test checks the email is sent once per booking

## Agent working
### Service list with prices (agent: claude-1)

## Review
### Contact page (agent: claude-2)
Changed: added the page and a map link.

## Done
- Home page with salon photos

Keep rules that apply to every card (how to run tests, what never to change) in your CLAUDE.md file instead, so you do not repeat them on each card.

Four ways to run an AI task board

Vibe Kanban

Vibe Kanban by BloopAI is an open-source board where each card can start a coding agent in its own branch, with a view of the changes. Its README lists ten supported agents, including Claude Code, Codex, Gemini CLI, GitHub Copilot and Qwen Code, and it runs with npx vibe-kanban. On April 10, 2026, bloop (the company behind it) announced it was shutting down and removed its online services. The app lives on as a community-maintained project, and local workspaces keep working. That makes it fine to try, but check it is still being updated before you trust it with long-term work. Our Vibe Kanban alternatives page covers where to move.

GitHub Projects

If your code is on GitHub, GitHub Projects gives you a board layout next to a table and a roadmap view. Cards can be issues, pull requests or draft issues, and built-in workflows can set fields when items are added or changed. If your team already tracks issues there, it adds nothing new to learn, but your agents do not see it unless you paste a card in or give them a tool that reads issues.

A plain markdown file

The TASKS.md file above. It costs nothing, it lives with the code, and every agent can read it. It has no drag and drop, and two agents editing the file at once can clash, so ask agents to change only their own card.

VibeCodingDeck's board per project

VibeCodingDeck gives every project its own to-do board and notes, right beside the grid of agents working on that project (up to 12 at a time). Each pane has a status light, so a card in "Agent working" and the agent's live state sit on the same screen. Agents can work on safe copies of the project, so two cards can run side by side and you merge the good results. For long plans, plans in batches run one agent per batch and wait for earlier batches to finish. You can also send users' crash reports from Sentry and GitHub issues straight to an agent. It runs on Windows 10/11 and Apple silicon Macs.

When the board stops matching reality

Boards for AI agents go wrong in a few predictable ways. Here is what each one looks like on the salon board, and the fix:

  • Review piles up. Agents finish faster than you can check. If Review holds more cards than you have agents, stop pulling new cards into Agent working until you catch up. Unchecked work tends to stack mistakes on top of each other.
  • A card bounces back and forth. If the "confirmation email" card has been sent back from Review three times, the card is unclear, and a fourth try will not help. Rewrite the goal and the "done when" list, or split it in two.
  • Cards sit in Agent working with nobody on them. An agent stopped, crashed or ran out of usage and the card never moved. Check each agent's status before you trust the column.
  • Ideas becomes a junk drawer. Once a week, delete ideas you no longer care about and turn the best two into Ready cards.
  • Two agents edit TASKS.md at the same time. One overwrites the other's update. Ask agents to change only their own card, and keep the file short by moving old Done cards to an archive file.

Card-writing rules for coding agents

  • One outcome per card. "Add a booking form and redo the menu" is two cards.
  • A "done when" list you can test by clicking. It turns Review into a quick check instead of guesswork.
  • Say where, and where not. "Booking page only, don't touch payments" saves you from surprise edits.
  • Small enough for one sitting. If a card would take an agent hours, split it.
  • Watch for shared files. Two Agent working cards that edit the same file is the most common cause of lost work.
  • Write follow-ups as new cards. Do not stretch a finished card to cover new ideas.

For bigger builds with many cards that depend on each other, read about AI agent orchestration, and see vibe coding best practices for habits that keep agent work tidy.

Task board questions

Does Claude Code have a kanban board?

+

Not a visual one. It has a task list inside a session, and with experimental agent teams the list is shared, with pending, in progress and completed states. Ctrl+T toggles it.

Is Vibe Kanban shutting down?

+

The company behind it, bloop, announced on April 10, 2026 that it was shutting down, and the GitHub README now opens with "Vibe Kanban is sunsetting". The app stays open source under the Apache-2.0 licence as a community-maintained project, and local workspaces still work. Plan a move for ongoing projects.

Can several agents work from the same board at once?

+

Yes, as long as their cards do not edit the same files. Give each agent its own copy of the project with a git worktree, and review each result before you merge it.

Should the board live inside my project?

+

A TASKS.md file inside the project is handy because agents can read it directly and git keeps its history. A separate board like GitHub Projects suits teams who already track issues there.

How big should a card for an AI agent be?

+

Small enough to finish in one session, with one clear outcome and a short list of checks. If you cannot write a "done when" list, the card is not ready yet.

Sources

Run up to 12 AI coding agents in one window.

Claude Code, Codex, Gemini, Grok and more, side by side. Windows and Mac. From $29 a month.

Get started

Keep reading