Multi-Agent Coordination Kit

A coordination architecture for multi-agent systems that solves three specific failure modes: agents that do not know what other agents are doing, agents that duplicate work on the same task, and agents that loop indefinitely when stuck.

The architecture uses an inbox/outbox model. Each agent reads from its own inbox folder, writes to other agents' inboxes, and never polls for responses. Messages have a defined format with sender, recipient, priority, task, and expected output.

Also covers: the escalation chain (agent to manager agent to human, with timeouts), the dead-letter queue pattern, and the TASKS.md shared file format for cross-agent task visibility.

What is included: the coordination protocol specification, message format templates, inbox/outbox setup guide, the TASKS.md shared file format, the escalation chain definition, and an annotated example tracing a task from request through completion.

Best for operators running three or more agents who are seeing coordination failures, duplicated work, or tasks falling through without resolution.