Are LLMs Truly Solving Software Problems — or Are Agents Doing It?

Two experiments were conducted to isolate the model and the agent factors, and contrast these two perspectives using SWE-bench as a controlled testbed.