Running agents

Start task sessions, monitor progress in real time, and run multiple agents in parallel.

Kanban gives each task its own isolated environment, so agents can work on separate tasks simultaneously without stepping on each other.

Starting a task

Click the play button on any card in the backlog column. Kanban handles everything that follows automatically.

1. Worktree created

   Kanban creates a git worktree branching from the task’s base ref. Each task gets its own working directory, so file changes from one task never interfere with another.

2. Gitignored files symlinked

   Directories like node_modules are symlinked from the main repo into the worktree. You get a fully functional project environment without running npm install (or its equivalent) for every parallel task.

3. Agent launched

   The selected CLI agent starts inside the worktree with the task’s prompt as its initial instruction. The card moves from backlog to in_progress.

Symlinks work well because you typically do not modify gitignored files during normal development. If your workflow involves changes to gitignored files (for example, writing to node_modules directly), do not use Kanban for those tasks.

Monitoring progress

Kanban uses runtime hooks to surface the agent’s latest activity on each card — the most recent tool call, file edit, or message. You can monitor dozens of parallel sessions at a glance from the board without opening each one individually.

To see the full agent terminal or chat interface, click on the card.

Stopping a task

Click the stop button on an in_progress card to halt the agent session. The card stays in the in_progress column so you can resume or inspect it later.

Running tasks in parallel

You can start as many tasks as you like at the same time. Because each task runs in its own worktree, agents work independently and do not produce merge conflicts with each other during execution. Conflicts are handled at review time when you choose to commit or open a PR.

Plan mode

When a task has plan mode enabled, the agent produces a written plan before making any file edits. This is useful for complex tasks where you want to review the agent’s approach before it starts writing code.

You set plan mode when creating or editing a task. You can also start a task in plan mode from the task detail view.

CLI alternative

Start a task from the command line with:

kanban task start --task-id <id>

The command starts the agent session and moves the card to in_progress. You can get a task’s ID from kanban task list.

# List all tasks and find the ID you need
kanban task list

# Start the task
kanban task start --task-id abc123