Agents based on large language models (LLMs) are increasingly entrusted with business decisions, yet their performance over long horizons is erratic. We examined whether coupling an agent with an external decision layer improves company management more than selecting a stronger model. On YC-Bench, a public benchmark in which an agent runs a start-up for one simulated year and is scored on final cash, eight LLMs were evaluated alone and with WhAI, a decision platform that estimates unobserved quantities from the history of the run and ranks candidate plans by Monte Carlo simulation of cash. In a paired test on four simulated worlds without published runs, all five tested models performed better with WhAI than alone in 48 of 48 pairs, with a mean gain of 1.72 million USD (95% CI 1.52–1.92; sign test p = 3.6 × 10⁻¹⁵). In additional analyses robust to clustering, the decision layer explained 72% of the variance in final cash, against 5% for the choice of model. On the published configuration, seven mid-size models with WhAI exceeded the best published model operating alone (3.45–3.82 versus 2.87 million USD) at up to one twentieth of its inference cost. In structured decision processes, the architecture surrounding a language model can matter more than the model itself.