Researchers demonstrated that optimizing the orchestration layer reduced the bill for agentic tasks by 33% to 61% and median latency by 44% across Claude Sonnet 4.6, Gemini 3.1, Flash 3.5, Qwen 3.6, GLM 5.1, and Palmyra X6. This optimization also decreased tokens per task by 38% while maintaining headline quality, showing that orchestration hygiene impacts cost more than model choice. The harness achieved these savings by implementing cache-shape discipline, structured compaction, and context offload, resulting in 99.9% of tokens served as cache reads.
Agentic AI compute costs are now primarily controlled by orchestration layer efficiency, not model choice, shifting the P&L lever from model providers to harness developers.