The current mainstream of enterprise AI development has devolved into "token maxing"—a flawed practice of buying system performance by mindlessly inflating token volume. According to research from the Writer, Inc. team, led by Muayyad Saeed Ali and Wassim Alshikh, the plummeting price per token merely masks Jevons Paradox: as model efficiency increases, businesses begin to consume resources exponentially. When moving from prototypes to industrial production, the hidden "reasoning tax"—multi-step Chains of Thought (CoT), bloated tool schemas, and quadratic context repetitions—creates a non-linear cost surge that can destroy a task's underlying value. It is now clear that base models are no longer the primary cost driver; rather, the money is being burned by the software surrounding them.
The Anatomy of Token Maxing
An agentic task, such as contract reconciliation, is never limited to a single model call. As noted by Writer, Inc., the process unfolds over dozens of iterations involving system prompts, retrieved data, and intermediate conclusions. In naive implementations, every previous step is recalculated from scratch, turning the budget into a sieve: the number of tokens per task grows faster than the utility of the result. Researchers have introduced the concept of a "harness"—an orchestration layer that manages the context window and determines the stopping point. Without strict discipline in this layer, companies face an avalanche of "failure-spend"—the cost of cyclical agent errors and unmanaged junk data extraction.
"Token maxing" is invisible in quality benchmarks, but it hits the wallet hard in cloud billing statements.
To prove the economic impact of architecture, the Writer, Inc. team conducted experiments across 22 evaluation tasks using six models, including Claude 3.5 Sonnet, Gemini 1.5 Pro, Gemini Flash 1.5, Qwen 2.5, GLM 4, and Palmyra X6. By keeping the models constant and replacing only the standard orchestration layer with the Writer Agent Harness, researchers isolated the economic effect of the software itself. The data confirmed it: the choice of harness affects task cost more significantly than switching from the cheapest model to the most expensive one.
The Leverage of the Harness
The methodology's results show that specialized orchestration allows performance to be decoupled from waste. Implementing the Writer Agent Harness reduced the average task cost by 41% (from $0.21 to $0.12) and median execution time by 44% (from 48 to 27 seconds). Crucially, this effect is universal: every tested model became 33–61% cheaper to operate. Researchers also identified a phenomenon called "harness leverage": a model's ability to extract quality from the orchestration architecture correlates almost perfectly with its baseline power (r=0.99).
Orchestration is the only component whose efficiency scales across all models a company uses today or adopts tomorrow.
For CTOs, this marks a paradigm shift: what matters is not which LLM you choose, but the "token economics" of your orchestration layer. Implementing control mechanisms—from caching discipline to failure-cost management—allows the number of completed tasks per million tokens to increase from 54.9 to 92.0. Instead of buying quality at the price of infinite context bloat, architects must focus on how the system sequences moves and delegates labor.
This research shifts the focus of AI ROI from pure benchmarks to architectural efficiency. The Total Cost of Ownership (TCO) for autonomous systems is now determined by orchestration design more than by an LLM provider's price list. While an optimized harness promises an 82% increase in quality-per-dollar, the reality is that current out-of-the-box agent solutions are economically unsustainable at industrial scale. The main risk remains the expansion of context windows: unless management mechanisms are baked into the harness itself, larger windows will only encourage more wasteful developer behavior in the hope of solving problems through brute computational force.