Question 1154

The Bill Multiplied

A single agent answering a research question used roughly 40k tokens. The team replaces it with an orchestrator that fans out to 8 subagents, each running its own multi-turn loop over a similar brief, then synthesizes. Token spend jumps far more than 8x. What best explains the jump?

A

Parallel requests are billed at a premium rate compared with issuing the very same calls sequentially

B

Each subagent resends its own growing transcript every turn, so spend scales with branches times turns

C

Fan-out forces every subagent to load the orchestrator's full transcript at the start of each of its turns

D

Output tokens dominate agent cost, and eight parallel branches produce eight full-length final answers