Researchers investigate whether large language models (LLMs) can automate software optimization, finding that LLM-based agents can achieve substantial performance improvements in scientific software, potentially making manual optimization obsolete for well-scoped, verifiable problems.