A new VS Code extension called Agent Kanban tackles one of the messier parts of coding with AI agents: keeping track of what you asked them to do, what they decided, and what actually got built.
The extension embeds a kanban board directly inside VS Code and uses markdown files as the backing store for each task. That last part is the interesting design choice. Instead of tasks living in some proprietary database or ephemeral chat history, each task is a .md file containing the plan, considerations, decisions, and actions taken. When your AI agent's context window fills up and it "forgets" earlier reasoning (a problem developers call context rot), the markdown file still has everything.
The workflow runs through structured commands - @kanban directives that move tasks through plan, todo, and implement stages. Rather than bundling its own AI agent, Agent Kanban plugs into whatever coding assistant you already use (Copilot, Cursor, Claude Code, etc.). That is a smart call. The AI coding agent space moves too fast for a task manager to bet on one provider.
Because tasks are plain markdown files checked into your repo, they work with Git out of the box. Your teammates can see task history, review AI-generated plans in pull requests, and pick up where an agent left off. That solves a real collaboration gap - right now, most AI coding sessions are invisible to the rest of the team.
Agent Kanban is free and available on the VS Code marketplace. It is a niche tool aimed squarely at developers who spend significant time directing AI agents through multi-step coding tasks. If your AI workflow is mostly "ask a question, get an answer," you probably do not need this. But if you are orchestrating agents across multiple files and features, having a persistent record outside the chat window fills a genuine gap.