“The best-performing AI agent on the SWE-bench benchmark achieves a success rate of approximately 62-63% on Python repositories.”