According to CoinMeta, as reported by AI, tests conducted by AI and agent on the einsia company's swe refactor showed that the success rate of migrating the entire codebase was only 5.4%. Out of 20 tasks, 13 were not completed at all. During the evaluation, AI and agent were given a time frame of 6 to 30 hours to complete their tasks, which included converting C code to rust, modifying maven to gradle, and migrating sqlite to wasi, among others. In the end, only 28 tasks were successfully completed out of 340 attempts, while 60 cases were found to have hidden bug. Among them, the configuration with claude and opus performed the best, successfully completing 5 tasks.