Commanding an AI Agent Army with Tickets and Kanban

The core

Past two or three agents, the management bottleneck isn't the model — it's tracking. The fix is simple: define work as tickets and reflect state on a kanban board. Exactly how people run teams.

Why tickets

A ticket fixes "what, why, and the done-condition." Hand an agent a ticket and instructions stop scattering — and you're left with a basis to look back at who decided what. That provenance is auditability.

The flow

1. Define: split work into tickets. Each ticket states outcome, scope, and done-condition.
2. Assign: the orchestrator routes each ticket to a fitting agent.
3. Track: state flows TODO → IN_PROGRESS → REVIEW → DONE on the board.
4. Record: each agent logs its progress on the ticket.

What the board gives you

• Visibility: see at a glance what's in progress and what's blocked.
• Accountability: each task's owner and state are clear.
• Retrospect: finished decisions and their rationale live on the ticket.

Where humans step in

You don't hand everything to automation. At REVIEW, a person checks the result, and decisions like merging stay with a human. The board makes those decision points explicit.

In Marblo

Marblo turns tasks into tickets, assigns them to agents, and tracks the whole army's progress on a kanban board in real time. A way to command many agents while staying in control — with tickets and the board at its center.