It's makers claim that an AI which can write the code to improve itself, more than doubled its performance In benchmarks like SWE-bench and Polyglot, outperforming hand-designed agents.