The 3× Token Bill We Didn’t See Coming

Transitioning to a multi-agent architecture unexpectedly tripled LLM token costs due to issues that were not immediately obvious. The root cause was not excessive agent calls, but rather expensive retry logic that would re-run entire upstream chains to rebuild context after a minor failure. The effective solution involved a three-part strategy: routing tasks to different models based on complexity, trimming unnecessary context between agent handoffs, and executing independent agent branches in parallel. These changes significantly reduced token consumption by preventing costly retry cascades and optimizing resource use for each step.