Coordination is the bottleneck for multi-agent, multi-user systems
Coordination is the bottleneck for multi-agent, multi-user systems. Everyone is bringing their own agents, and those agents are going to spar and argue with each other. If you want a system that's efficient, aligned, and focused on the big picture, you probably need some form of coordination.
In general, you want agents to be independent and to move fast. This is related to my thoughts on engineering orgs where each small team owns a microservice, with some coordination at the platform level. Pure decentralization with no structure is chaotic.
So what should that coordination look like? Maybe it's a smart layer over a shared brain or system of record, coordinating things like a shared calendar. Or maybe it's central, top-down coordination, like a CEO.
But a CEO doesn't tell everyone what to do, because that doesn't scale. A CEO sets the direction of the company, and conflicts get handled at a lower level. When you disagree with a coworker, you don't escalate right away; you talk it through and resolve it with your coworker. Escalation is a last resort. That's how human organizations scale.
I wonder if things are different in an agentic world. Agents might work fast enough that a central coordinator stops being a bottleneck, or the coordinator could be infinitely replicated. Then the hard part becomes the system of record. At a human organization's pace, it doesn't need to be very resilient. At agent speed, it probably does.