Coordination Is the New AI Bottleneck

For three years the enterprise AI question has been a capability question. How good is the model. How does it score. How much better is this quarter's release than last quarter's. That question mattered when the models were the limiting factor. Increasingly, inside organizations actually deploying this technology, they are not.

The previous essay described the agentic enterprise as a system of delegated machine action, and Arc I ended on delegation networks as the new unit of scale. Put those together and a different bottleneck comes into view. When work is delegated across many actors, some human and some machine, the thing that determines whether the system performs is not the intelligence of any single actor. It is how well the whole is coordinated. And coordination is exactly the capability that is not advancing on the same curve as the models.

This is the gap that should worry executives more than any benchmark. Model capability is improving fast and will keep improving. The organization's ability to coordinate people, agents, tools, context, and decisions is improving slowly, because it is an operating-model problem, not a technology you can buy. The result is a widening mismatch: more capable actors dropped into structures that cannot coordinate them. A brilliant model in an incoherent system produces fast, confident, poorly integrated work, which is not an improvement over slow work. It is a new failure mode.

Coordination cost grows with the number of actors

There is a reason coordination becomes the binding constraint, and it is structural rather than technological. Every actor you add to a system adds not just capacity but connections, handoffs, shared context, and points where something can be dropped. A single expert doing a piece of work coordinates with no one. Two people coordinate across one relationship. A network of a dozen humans and agents coordinates across a web of them, and the cost of keeping that web coherent grows faster than the headcount does. Adding a smarter actor to a poorly coordinated network does not fix the network. It just lets the incoherence happen at higher quality and greater speed.

This is where the systems fail in practice, and they do not fail at the model. They fail at the seams. A handoff drops the context that the next actor needed. An agent proceeds on a stale assumption because no one propagated the update. Authority is ambiguous, so two actors both act, or neither does. Evidence produced early in the chain is lost by the time a human has to sign off. None of these are intelligence failures. They are coordination failures, and no amount of additional model capability removes them, because they live in the connections between actors, not inside any one of them.

The next AI bottleneck is not intelligence. It is coordination.

What coordination actually involves

It is easy to say coordination and mean nothing precise, so let me define it as work rather than as a value. Coordinating delegated work means decomposing a goal into units that can actually be delegated. Routing each unit to the human or machine actor suited to it. Sharing the context each actor needs, and only what it needs. Reviewing outputs where judgment is required. Escalating when a unit exceeds the authority it was given. Preserving evidence so conclusions can be traced. And carrying memory across the whole thing so the system does not relearn what it already knew. That is the actual content of coordination, and none of it is soft.

Two things follow from that list. First, this is not collaboration in the cultural sense of people getting along with better software. It is an engineering and governance discipline, closer to how a controls engineer thinks about a plant than to how a team thinks about teamwork. Second, it is not project management either. Project management sequences tasks among actors who carry their own judgment and accountability. Coordination in an agentic system has to supply the judgment placement, the authority boundaries, and the evidence flow that human actors used to bring implicitly, because some of the actors now bring none of it on their own.

Span of control, again

One of the oldest ideas in organizational design returns here in a new form. Span of control, the limit on how many subordinates one supervisor can direct effectively, was a fact of human management long before it was written down. Delegation across human-machine networks does not abolish that limit. It relocates it. The question becomes how much delegated work, across how many actors, an orchestrating layer can coordinate before coordination itself degrades.

Systems that ignore this over-delegate. They spawn agents and hand off work faster than the coordinating layer can keep coherent, and the network quietly loses the thread. The disciplines that hold it together are the ones the models will not provide for you: bounded delegation depth so authority chains do not run away, clear escalation so decisions that exceed authority surface rather than proceed, and evidence trails so the whole can be reconstructed and trusted. The failure of a poorly coordinated agent network is not that it is dumb. It is that it is confidently, rapidly, and untraceably wrong, which is worse.

Where the advantage moves

If coordination is the bottleneck, then the advantage moves to the organizations that build coordination infrastructure rather than the ones that simply acquire the strongest models. Models are becoming a commodity input; any serious competitor can buy comparable capability. Coordination is not a commodity. It is built, it is specific to the organization, and it compounds, because a system that coordinates well gets more out of every model it plugs in and keeps that advantage as the models change underneath it.

That is the investment thesis this arc points at. The enterprises that win will not be the ones with privileged access to the best model. They will be the ones that treated the coordination of human and machine work as strategic infrastructure and built it deliberately: decomposition, routing, context, review, escalation, evidence, and memory as designed capabilities rather than emergent habits. The last item on that list is doing more work than its position suggests. Coordination depends on shared, durable context, and most enterprises do not have it, which is the subject of the next essay: the enterprise memory problem.