The term 'fabric' suggests the woven interconnect between agents — a substrate that carries messages, events, and state without any single agent needing to coordinate with others directly. Unlike a simple event bus (which handles publish/subscribe), a fabric typically includes additional coordination primitives: inbox messages (direct agent-to-agent communication), work-claiming (exclusive reservation of a task by one agent), and persistent storage of all events for audit and replay.
A fabric makes the coordination topology of a multi-agent system observable: you can query which events have been published, which agents have claimed which tasks, and which messages are waiting in which agent inboxes. This observability is critical for debugging coordination failures and for auditing agent decisions.
Fabrics can be centralized (all agents share one fabric instance) or federated (separate fabric instances per team or project, with cross-fabric routing for events that span organizational boundaries). The tradeoff is consistency versus isolation: a centralized fabric provides a single source of truth but is a single point of failure.