Enterprises using AI face a cost problem: throwing more computing power at tasks works in testing but becomes too expensive in real use. A new study from AI company Writer offers a practical fix. Instead of changing the underlying AI model, the researchers optimized the 'harness' — the software layer that connects the model to applications. By improving how the harness manages prompts, caches data, and delegates tasks, they cut token usage by 38% and reduced the cost per successful task by up to 61%. Task success rates actually improved slightly, from 78% to 81%. The key changes include separating stable instructions from changing queries to reuse cached data, moving old conversation history out of the main prompt, and using small sub-agents for specific jobs. These techniques are under the developer's control and don't require retraining the AI model. The study tested six different models from various vendors, including Writer's own Palmyra X6, on 22 enterprise tasks. The approach works best with stronger models; smaller ones struggled with sub-agent delegation. For companies scaling AI to millions of requests daily, these optimizations can lead to major savings. The researchers warn that developers should avoid 'tokenmaxxing' — relying on huge context windows and brute-force retries. Instead, they recommend structured prompts, context offloading, and hard budget limits on each task. As AI models improve, the harness will focus more on enforcing company policies like budgets and permissions rather than compensating for model weaknesses.