Skip to content

What is Agentic Kanban?

Agentic Kanban is a VS Code extension that gives you and a coding agent a shared Kanban board. Tasks live as Markdown files your team can read, edit, and review in Git — no opaque database, no context drift between sessions.

Pick a profile that fits your work style:

  • Litebacklog → in-progress → done. Fast path for bug fixes and small changes.
  • Standardbacklog → planning → in-progress → review → done. Full spec-driven delivery with plan approval and implementation review.

A visual board — cards you can drag between lanes, with priority, labels, and dependency indicators visible at a glance.

@kanban chat commands — create tasks, attach specs, create worktrees, run diagnostics, and drive lane sweeps from the VS Code Chat panel.

Spec artifacts per task/spec auth scaffolds proposal.md, design.md, tasks.md, and a shared capability spec under .agentkanban/changes/. The agent works from these files, not from stale chat memory.

Loop-until-dry automation/loop repeatedly advances ready tasks through lanes until the lane is dry. Dependency-aware: task B stays parked until task A reaches done.

Persistent context — a managed section in AGENTS.md and .agentkanban/INSTRUCTION.md keep the workflow rules and active task visible to any agent on every turn.


Every card on the board is a Markdown file under .agentkanban/tasks/. Example:

---
title: Add OAuth2 login
lane: planning
priority: high
labels:
- backend
dependsOn:
- establish-auth-storage
change: .agentkanban/changes/add-oauth2-login
---
## Conversation
### user
Plan the OAuth2 implementation.
### agent
Starting with auth boundary mapping...

The board reads these files to render columns. Any card move on the board modifies the lane: in frontmatter. This makes your workflow fully auditable and Git-trackable.

A task always sits in a known lane. Lanes are fixed per profile — agents cannot invent new lane names or skip steps:

  • Lite: backlog → in-progress → done
  • Standard: backlog → planning → in-progress → review → done

This keeps the workflow coherent across long agent sessions where chat context has long since rolled off.

The extension enforces exactly two human checkpoints in Standard profile:

  • Plan Approval: moving a task from planning to in-progress is your sign-off that the plan is ready.
  • Completion Gate: moving from review to done is your sign-off that the implementation is accepted.

Everything in between (implementing, verifying, checking off tasks.md) the agent handles autonomously. /loop is specifically designed to refuse crossing either gate — it will park tasks that would need a human decision rather than skipping ahead.

Tasks declare dependencies in frontmatter:

dependsOn:
- establish-auth-storage
labels:
- blocked-by:establish-auth-storage

When you run @kanban /loop, it checks each task’s dependsOn list. A task is only ready when every dependency is in done. Independent tasks can be processed in parallel; dependent chains stay strictly ordered. The loop runs multiple passes until nothing more can advance.