Coordination is an organizational problem

When people imagine teams of AI agents, they tend to imagine a relay race: agent one finishes and hands off to agent two, work flowing cleanly down a pipeline that someone drew in advance.

Put multiple agents on a real problem and what you get looks less like a relay race and more like a badly run company. Two agents do the same work twice. An agent waits forever for an answer that another agent does not know it owes. Stale information gets acted on with complete confidence. A mistake made by one agent passes through three others before anything, human or machine, notices that the original premise was wrong.

I find this genuinely encouraging, and I want to explain why.

The failures of multi-agent systems are not exotic. They are the ordinary failures of organizations: unclear ownership, missing protocols, no shared source of truth, no escalation path, accountability that pools nowhere. Anyone who has worked inside a company of more than ten people has lived through every one of these. Which means the solutions are not waiting to be discovered. They are waiting to be translated.

Humans are individually unreliable, and human organizations are, at their best, astonishingly reliable. Aviation, surgery, power grids: these run on people who forget things, get tired, and make mistakes, wrapped in structure that catches what individuals drop. Checklists, handoffs, sign-offs, audits, post-mortems, separation of duties. Ten thousand years of learning how to build dependable institutions out of undependable minds.

That body of knowledge is the most underused resource in artificial intelligence today.

We see it directly in our own systems. Give agents explicit roles instead of letting them negotiate who does what, and coordination failures drop. Route agent-to-agent communication through structured handoffs (here is what I did, here is what I could not do, here is what you need to know) instead of free-form conversation, and error propagation drops. Make one agent accountable for an outcome with others as contributors, instead of everyone being vaguely responsible, and the outcome actually gets owned. None of this is machine learning. All of it is management.

There is a real research program here, and it is not the one usually described. The usual framing is communication protocols, negotiation, swarm behavior, emergent cooperation. Ours is closer to: what does an organization chart for machines look like? What is the minimum institutional structure that lets a hundred fallible agents produce one reliable result? Which human institutions translate directly, which need rework, and which exist only to compensate for limitations machines do not have?

That last question is the interesting one. Some institutional structure exists purely because of human constraints. Meetings exist because human memory does not synchronize. Layers of management exist partly because human attention does not scale. Some of the org chart will dissolve when the minds inside it can share state directly. The question is which parts, and the only way to find out is to run the experiments.

Multi-agent coordination is the laboratory’s second major direction, after individual autonomy and ahead of self-improvement. Not because it is fashionable. Because every outcome that matters in the world is produced by more than one mind working with others, and there is no reason to expect machine work to be different.