Gareth Brown, a UK-based developer at AppSoftware, has released VS Code Agent Kanban — an open-source extension targeting a problem most AI coding users know well: context rot. Long sessions with GitHub Copilot or similar tools tend to degrade as the agent loses track of earlier decisions, trade-off discussions slip out of scope, and behavior drifts. Brown's fix is to embed a Kanban board directly inside VS Code, backed by plain Markdown files that record the full history of an AI-assisted task without leaving the IDE.
The architecture is deliberately minimal. Every task lives as a .md file with YAML frontmatter inside a .agentkanban/tasks/ directory, using [user] and [agent] markers to log the planning conversation as it unfolds. Because the format is just text, the entire folder can be committed to Git — giving teams a shared, auditable, and diffable record of what the AI was asked to build and why. A @kanban GitHub Copilot Chat participant exposes a structured plan/todo/implement workflow, prompting the developer to confirm intent at each phase before handing off to Copilot's native agent mode for execution. A /refresh command re-injects full task context whenever the agent begins to drift during a long session.
Brown ditched an early attempt at a custom LLM execution loop after running into the limits of what the VS Code extension API can reasonably handle. The current version delegates code generation, tool calls, diffs, and terminal operations entirely to Copilot's agent machinery — no bundled model, no competing tool runner. Existing instruction files like AGENTS.md and CLAUDE.md continue to work without modification. A post-launch update added Git worktree support: each task can spin up an isolated branch and working directory, so the agent can make sweeping changes without disturbing the main workspace.
VS Code Agent Kanban is on the VS Code Marketplace now, with source at appsoftwareltd/vscode-agent-kanban. For teams already on Copilot, the practical gain is straightforward: task history accumulates in version control, context survives session restarts, and anyone on the project can see exactly what the agent was asked to do and why. There's no new server to run, no additional subscription, and no migration away from existing Copilot workflows.