This talk proposes a shift toward 'evolvable compilers' by using AlphaEvolve, a Gemini-powered agent, to automatically generate and refine compiler optimization heuristics for LLVM and XLA. By applying this method to challenges like function inlining and register allocation in LLVM, the project aims to reduce manual tuning efforts and create more performant, concise optimization policies. The results demonstrate that these automatically generated heuristics can outperform human-designed ones, with plans to upstream the solutions to the LLVM and XLA communities.