What coordination overhead does a multi-agent system introduce that a single-agent system avoids?

A **single-agent system** is one model in one loop. It makes every decision, holds the whole task in one context, and never has to talk to anyone. A multi-agent system splits the work across several agents, usually with one orchestrator that hands sub-tasks to specialists. That split is not free. Three new kinds of work appear. First, routing: something has to decide which agent handles which sub-task, and it can route wrong. Second, shared state: if two agents touch the same document or record, you have to keep their views consistent. Third, messaging: agents must pass results to each other, which adds latency and new ways to fail. A single agent skips all three. It never routes because there is only one worker. It never synchronises because there is one context. It never sends a message because there is no one to send to. So when is the overhead worth it? Only when the task truly breaks into independent pieces that can run at the same time, or when a real specialist clearly beats a generalist. For most tasks, one well-equipped agent is simpler and cheaper.

A multi-agent system introduces three forms of coordination overhead that a single agent never pays. The first is **routing**: an orchestrator must decide which specialist handles each sub-task. That decision point can misroute, and it adds an extra model call. The second is shared-state synchronisation. When agents work on related sub-tasks they often need common state, like a document or an intermediate result. Keeping that consistent without stale reads or race conditions adds real architectural complexity. The third is inter-agent messaging. Agents must pass results to each other or back to the orchestrator. Every hop adds latency and a new failure surface, such as a dropped message or a schema mismatch between two agents. A single agent avoids all three because there is one loop, one context, and one decision maker. The overhead is worth paying only when sub-tasks are genuinely independent and parallelisable, or when a specialist clearly beats a generalist. Parallel research over many sources is a good fit. A tightly coupled, sequential task is not.