A recent paper, 'The Harness Effect,' demonstrates that the cost of enterprise AI agents is predominantly determined by the orchestration layer, not the unit price of the underlying model. By optimizing how context is organized, history is compressed, and tool calls are managed, the study achieved a 38% reduction in token consumption per task, leading to a 33-61% decrease in overall costs and a 68% improvement in cost per million tokens (CPM). This finding is crucial for engineering leaders and architects, as it shifts the focus from selecting cheaper models to designing more efficient agent architectures. The practical implications are significant: teams can achieve substantial cost savings by refining their orchestration strategies rather than solely relying on model price reductions.
Enterprise agent costs are driven by orchestration design, not model price. Optimizing context and tool calls can reduce costs by 33-61%.