Coordination Layers

Multi-agent systems operate on a spectrum of coordination patterns. Sequential dependencies favor supervisors, domain-separable work favors hierarchy, and exploratory work favors peer or swarm arrangements. Pattern choice follows the work, not preference.

Centralized control is one point on a spectrum. The alternatives relax it in different directions, changing what you can see and what you have to govern. The platform layer matters as much as the pattern, because it determines whether state stays visible once agents cross framework boundaries.

Swarm patterns show many lightweight agents acting on shared signals with minimal explicit routing. This works well for broad search and triage tasks where coordination overhead outweighs the benefits of explicit communication.

The operational layer beneath coordination often matters more than the pattern itself. When agents move between frameworks, state visibility becomes the critical constraint. A supervisor pattern fails when the supervisor cannot see the state of tools running in other environments.

Production systems must account for both the coordination pattern and the underlying substrate. An elegant decentralized design breaks when agents cannot reliably discover each other's capabilities. The protocol determines whether the abstraction remains consistent at scale.