Multi-agent systems multiply cost
Every additional agent in a system is another full round of model reasoning, and that means multi-agent systems multiply the latency and token cost of a single agent rather than simply adding a small overhead. A supervisor pattern with three workers isn't running one model call, it's running at least four, the supervisor's planning call plus each worker's own reasoning, and often more once you add message formatting, retries, and a final assembly step. A sequential pipeline pays a similar cost, with each stage adding its own latency on top of the last.
This cost is easy to underestimate when a demo only runs once, but it adds up quickly in production, both in dollars and in the time a user waits for a final answer. A five-agent system that takes 30 seconds and costs ten times as much as a single well-prompted agent is only worth it if it produces a meaningfully better or more reliable result.