The Harness Effect: How Orchestration Beats Token Maxing on Your AI Bill

A controlled swap across six frontier models shows the orchestration layer, not the model, determines the cost of agentic AI, according to a practitioner who has built agent systems for years. The author argues that an agent request unfolds into multiple turns—system prompts, tool schemas, retrieval payloads, model reasoning—and that the software wrapped around the model drives the invoice gap more than swapping the priciest model for the cheapest one.

Member-only story The Harness Effect: How Orchestration Beats Token Maxing on Your AI Bill A controlled swap on six frontier models shows the orchestration layer, not the model, sets the cost of agentic AI. Have you ever watched two teams run the exact same frontier model and walk away with invoices that aren’t even close? Same Claude. Same Gemini. Same prompts on paper. One team burns a fortune. The other doesn’t. For years I chalked that gap up to prompt skill or luck, or somebody leaving a debug loop running over a weekend. I was wrong about where it comes from. The gap doesn’t live in the model. It lives in the software wrapped around it. I’ve spent the last couple of years building agent systems for a living, and I read AI infrastructure research the way some people read box scores. So a claim that the orchestration layer moves your cost more than swapping your priciest model for your cheapest one stops me cold. Let me show you why I now think it holds up. The word most people skip past An agent request is not about one LLM call. Ask it to “ reconcile these two contracts and draft the redline ,” and that request unfolds into a dozen turns. System prompt, tool schemas, retrieval payloads, model reasoning, tool…