Agent scaffolds are increasingly designed to offload small decisions—whether to stop, verify, or change approach—to cheaper models. A new preprint on arXiv asks whether a typed decision model, Jev 1.13, consulted only at deterministic decision points, can improve the performance of a Claude-based agent.
The paper's title, "Good Judgment, Bad Advice," summarizes the central finding: the decision-model controller raises agent cost without improving success. While the abstract is truncated, the stated conclusion is unambiguous about the trade-off.
This is a single preprint, so the results are preliminary and not yet peer-reviewed. Still, the work highlights a potential pitfall in a common design pattern for agent scaffolds, suggesting that adding a decision layer may not be worth the extra expense.