Why coding agents need more than a prompt
A coding agent can edit a repository, run tests, and prepare a commit. The difficult part begins when several tasks or agents run at once: prompts become detached from their results, terminals disappear, work overlaps, and reviewers cannot quickly tell what changed or why.
Agent Kanban PM provides a local control plane around that work. A card records the objective and state. A session records the execution. A worktree can isolate the files. Review evidence ties the result back to the task. The board gives both the human operator and connected agents one shared view of progress.
The plan, run, review loop
1. Plan work as durable tasks
Create a card with a concrete objective, acceptance criteria, dependencies, and relevant repository context. Agents can read and update the same structured task data through the UI, CLI, REST API, or MCP interface.
2. Run in an isolated workspace
Assign a task to an installed coding-agent CLI and launch it from the board. The runtime tracks the process and terminal output. Git worktree support lets concurrent jobs operate without competing for the same checkout, while explicit leases prevent two workers from claiming the same task accidentally.
3. Review evidence, not status theater
Moving a card to review is not a claim that the work is done. The task can carry its commits, changed files, test results, agent output, and approval state. A reviewer sees what happened and can approve, return, or recover the work with its context intact.
Designed for failure and resumption
Long-running agent work is not always clean. Terminals disconnect, processes exit, tools time out, and a task may need another pass. Agent Kanban PM treats the task record as durable and the process as replaceable. Recovery controls and event history make it possible to resume deliberately instead of reconstructing the run from chat logs.
Local-first by design
The application runs on your machine and stores its operational data locally. It does not require a hosted workspace or send task history to a project-management cloud. The coding agents you choose may have their own network and account requirements, but the Kanban layer itself does not.
Agent Kanban PM is MIT licensed and free forever. There is no subscription, premium edition, account, or usage meter.
Where it fits
This approach works well for a developer or trusted local team coordinating implementation, tests, reviews, migrations, and release chores across one or more repositories. It is intentionally not a hosted multi-tenant issue tracker. Its strength is close supervision of local coding-agent execution and exact Git handoffs.
Use the tools you already have
The project is agent-agnostic. Claude Code, Codex, OpenCode, Aider, Antigravity CLI, and custom shell adapters can all participate. See the coding-agent integrations, or follow the getting-started guide to launch your first board.