InferenceBench: agents hit 8.08x over naive PyTorch, a plain hyperparameter sweep hits 11.53x.

Give an agent one H100, a target LLM, and two hours of wall clock to deploy and optimize an OpenAI-compatible inference server across four scenarios (arXiv 2607.20468). Across 15 frontier agent configurations, agents reach up to 8.08x and often match default vLLM at 4.05x. A simple hyperparameter search under the same budget hits 11.53x. Trajectory analysis pins the gap on exploration, not knowledge: agents name many optimization techniques, converge on one framework, test a few configs, then burn the rest of the budget re-measuring and repairing. That's a specific, fixable behavior pattern and I'd bet it shows up in your agents too.