When multi-agent is worth it

The honest conclusion from these articles: multi-agent systems shine for high-value tasks that involve heavy parallelization, information that exceeds a single context window, and complex tool interfaces. They are a poor fit for tightly interdependent work like coding, where shared context matters more than breadth. And they are expensive, around 15 times the tokens of a chat, so the task has to be valuable enough to justify the cost.