We evaluate a simple inference‑time scaling technique called turn‑level best‑of‑N sampling (\method) for large language model (LLM) agents tackling complex, long‑horizon tasks. The approach uses a generator‑selector framework with two LLMs: at each turn the generator produces $N$ candidate actions, and the selector picks one to execute in the environment. This is related to the guided‑search method of \cite{zainullina2025guided}, but unlike that work we test whether off‑the‑shelf, open‑weight LLMs can directly compare candidate actions without additional training. Experiments on several long‑horizon terminal tasks show that \method improves accuracy by roughly 5%–10% and reduces variance across repeated runs. Trajectory analysis reveals a shift toward desirable behaviours such as fewer attempts to solve the task, better instruction following, and more systematic debugging and testing.\n\nReview