Agent orchestrators and agent coordinators are not the same layer

Agent orchestrators and agent coordinators are distinct layers within a system that manages multiple agents. Although the terms are often used interchangeably, they have different roles, authority, failure modes, and reasons for existing. This distinction is crucial to understand in order to effectively manage agent workloads.

An orchestrator is a crucial component that controls the flow of agent operations. It owns the run loop, which means it is responsible for spawning sessions, assigning each session to a specific isolated environment, determining the order in which tasks should be executed, restarting any failed sessions, and routing prompts and results. The orchestrator essentially owns the entire control path, allowing it to start or stop agents, manage sessions, and provide a single, consolidated view of the system's state.

On the other hand, a coordinator sits alongside the control path rather than within it. It does not start or stop agents, nor does it have authority to halt a keyboard. Instead, its primary role is to verify the compatibility of concurrent agent operations. Coordinators receive declarations about what each agent intends to do, such as modifying specific parts of the codebase.

Rather than executing tasks directly, coordinators focus on comparing these declarations to ensure that the intended changes can coexist without conflicts. They provide evidentiary findings and gatekeeping mechanisms to help prevent potential issues, such as declaring that one intent will replace a symbol while another will extend it, leading to a destructive_vs_additive conflict.