Introducing AgentKanban
AgentKanban is a context and workflow management tool designed for 'Human in the Loop' agentic software development tasks. This is achieved via a web application focused around Kanban Board task management and a VS Code extension that integrates with Github Copilot (with other harnesses in development) for context loading and capture.
AgentKanban automatically captures user / agent chat turns to the remote task on agentkanban.io, where you can use AgentKanban to resume and manage context in future chat sessions, or simply use as a record of decisions and task progress.
The AgentKanban MCP server and APIs can be used by agents with an API key to access broader context around your project where instructed.
Note: AgentKanban is only active when working with a selected board and task. You can also opt not to work with an AgentKanban task and use your harness as normal (e.g. for small / trivial changes).
If you are already familiar with AgentKanban and just want to know how to get set up, go to Getting Started.
AgentKanban boards
Tasks are created and edited collaboratively on AgentKanban task boards:
Kanban Board:
Task:
AgentKanban VS Code extension
In VS Code, the extension (VS Code Marketplace Open VSX Registry) connects to your boards using an API key. Task context is loaded into VS Copilot chant using Resume buttons. Optionally use the PLAN TODO IMPLEMENT workflow and Git worktrees. User / Agent turns are captured on the task for future context management (pin, fork, split delete etc).
AgentKanban Sidebar:
AgentKanban supported CoPilot Chat:
Why AgentKanban
Agentic software development produces the best results when the workflow is structured and you have control over the context.
Workflow
The optimal workflow for large tasks when working with AI agents includes:
Planning
- Collaborative real-time multiplayer board management and task creation / editing for use in teams.
- A thorough planning phase clearly defined with an opportunity for the agent to ask clarification questions.
TODOs
- A 'TODO' phase where the agent breaks the task down into iterations and TODO items.
- Task assets management and storage.
Implementation
- The agent implements the plan according to the plan and TODOs, keeping TODOs up to date as work progresses.
How AgentKanban supports this workflow
General
- Bring your own agent harness (The VS Code extension for Github Copilot). Use the harness that works for you. Agent conversations happen in your harness chat as normal.
- The extension auto-configures MCP tools so your agent can read and update tasks, TODOs and board memory directly.
- Automatic deterministic capture of chat turns to the remote task. No manual copy/paste needed.
- Persistence of context for auditing and resuming work on tasks over multiple sessions.
- Provide extension access to boards via MCP and API endpoints, secured with API keys that facilitate granular, cumulative permission levels.
- Security: TOTP 2FA, backup codes, passkeys (WebAuthn / Touch ID / Windows Hello / security keys) supported.
- Webhooks for integration with your existing tooling and workflows (HMAC-SHA256 signed payloads).
Planning
- Manage user sharing, assignment, task priority, lane and ordering for board tasks to help you prioritise and schedule work.
- Battle tested agent instruction based on Matt Pockock's "Grill Me" skill that encourages a robust planning stage for non trivial tasks.
TODOs
- AgentKanban instructs your agent to manage TODOs on the board tasks. You can view and collaborate on TODO tasks before or during agentic workflows begin.
Implementation
- Your agent implements the agreed plan, keeping TODOs up to date as work progresses.
Context Management
AGENTS.md instructions in the task board settings
Write AGENTS.md instructions in the task board settings and they will be injected into an AgentKanban managed AGENTS.md section where ever the board and task is selected in the agent harness extension. This means you write AGENTS.md once per task board (which usually aligns with a project) and all users of the board benefit from the shared instruction. Your existing AGENTS.md instructions are respected and will remain outside of the AgentKanban managed section.
Capture, Resume, Fork, Split, Pin and Delete agent context
AgentKanban captures conversational user / agent context without the bloat (tool use etc) allowing you to clear context and resume in new chat sessions to keep the context fresh when working on long running tasks or picking up tasks for re-work. See Agent Context Management and Turn Capture for more information.
Memory capture on the task board
AgentKanban will prompt your agent to record information as memory as it discovers durable technical context, conventions, or decisions that should be preserved beyond the current task.
Get Started
Ready to get started? See Getting Started for setup instructions.