Writer's AI harness cuts token spend nearly 40% — without sacrificing accuracy

Writer researchers found that improving the orchestration layer around a foundation model can reduce tokens per task by nearly 40% and cost per successful task by up to 61%. The study argues that teams can maintain accuracy without changing or fine-tuning the underlying model by optimizing the AI harness instead.

The findings challenge the common practice of “tokenmaxxing,” where developers rely on large context windows and repeated retries rather than better workflow design. As enterprise AI moves from experiments to production, cost-efficient orchestration may become as important as model selection for achieving sustainable ROI.