The numbers worth remembering

The comparison article gives a few rough figures that are useful anchors:

• A single agent uses roughly 4 times the tokens of a plain chat interaction.
• A multi-agent system uses roughly 15 times the tokens of a plain chat.
• Single agents are described as high and predictable in reliability; multi-agent systems as lower, with emergent (hard to predict) behaviors.

These are not precise laws, but they capture the tradeoff: multi-agent buys you breadth and speed at a large cost in tokens, reliability, and debugging difficulty.