Your Multi-Agent Pipeline Isn't Slow Because of the Model

Short summary

A developer running a custom Claude Code orchestrator discovered that multi-agent pipeline latency comes from excessive cold-start dispatches, not model speed. By merging redundant roles (5 dispatches down to 3 per work unit), parallelizing independent siblings, and short-circuiting trivial cases, they achieved a 40% cut in dispatches with zero loss of rigor. The article also details non-obvious prompt caching economics: cache reads at 0.1x base price, minimum cacheable prefix thresholds, and the importance of verifying cache behavior empirically.

• •Dispatch count—not model speed—is the primary bottleneck in multi-agent pipelines; each cold start re-pays full context reconstruction cost
• •Merging sibling roles that share context (researcher+planner, reviewer+auditor) cut dispatches 40% with no rigor loss
• •Prompt caching details matter: cache reads are 0.1x base price but minimum cacheable prefix and cache_control behavior must be verified empirically

Generated with AI, which can make mistakes.