Why Kanban fits agent work
Chat interfaces are fine when you have one thing to do. Once you have eight Claude Code sessions in flight, you're not having a conversation anymore — you're running a small team. Teams need a board.
Kanban is the format we already trust for parallel work. It gives you WIP limits, status at a glance, and a place for things to live between “started” and “merged.” Everything an orchestrator needs to show you, Kanban already knows how to display.
Claude Code's new desktop app puts parallel sessions in a sidebar. That works until it doesn't. A board scales to the number of tasks you actually have in a sprint, not the number that fit in a sidebar.
The five columns
Backlog. Everything the sprint generator produced, plus anything you drag in. Priorities and story points visible at a glance.
In Progress. Tasks currently running. Live model, live status, live terminal output one click away.
QA. The orchestrator ran the task and checked the validation criteria. If something failed, it lands here with the error attached.
Done. Completed work in your actual repo. Review the changes when you get back.
Failed. Retries exhausted. The full run history is preserved — you decide what to do next.
A session sidebar asks “what are you working on right now?”. A Kanban board answers “what's the state of the whole sprint?”.
What's on each card
Task description + validation criteria — the agent reads both. It knows when it's done because you told it what done means.
Model assignment — Haiku, Sonnet, Opus, or Value Mode (auto-route by complexity). Codex and Kimi too.
Priority + story points — the orchestrator uses these to pick run order.
Run history — every attempt, every error, every retry, with full terminal output preserved.
Keep reading
Kanban for AI coding agents — the full feature page.
DevboardAI vs Claude Code Desktop — the session sidebar vs the board.