What is the primary coordination cost that multi-agent systems pay that single-agent systems avoid?

A **multi-agent system** is several LLM agents working together on one task, instead of a single agent doing everything itself. It sounds more powerful, but it comes with a hidden bill called coordination cost. That cost has three parts. First, routing: something has to decide which agent handles which piece of work. A single agent never makes that decision because it does every piece itself. Second, shared state: when agents work together, they each hold a partial view of the task, and those views have to be kept consistent. Third, message passing: agents hand results to each other, and every handoff adds delay and a chance for the message to be lost, garbled, or misunderstood. A single agent running step by step avoids all three. It keeps one consistent memory of everything it has done, so there is nothing to route and nothing to synchronise. The takeaway is that multi-agent is not the default upgrade. You only reach for it when the task has real parallel branches or needs genuinely different specialists. Otherwise the coordination tax makes it slower and more fragile than one good agent.

The primary cost of going multi-agent is **coordination overhead**, and it shows up in three places. Routing decides which agent handles which sub-task. Shared-state synchronisation keeps each agent's view of the task consistent. Message passing moves results between agents, adding latency and delivery failures. A single agent avoids all three. It runs sequentially with one consistent transcript, so there is no routing decision, no state to reconcile, and no handoff to drop. The deeper problem is error propagation. In a single agent, a mistake stays local and the model can often see and correct it on the next turn. In a multi-agent pipeline, one agent's wrong output becomes another agent's trusted input. Errors compound across handoffs, and per-step success rates multiply. Ten handoffs at ninety percent each leave you near thirty-five percent. So multi-agent is not a free upgrade. It pays off only when the task has genuine parallelism, several branches that can run at once, or genuine specialisation, sub-tasks that need distinctly different tools or prompts. For most tasks, one capable agent in a loop is faster, cheaper, and easier to debug.