AI agents that do cosmology research without being asked twice

CMBEvolve uses large language model-guided code evolution and tree search to iteratively improve performance on benchmark tasks, demonstrated here on out-of-distribution detection in weak-lensing maps. CosmoEvolve runs a virtual multi-agent research lab capable of open-ended workflows, autonomously analyzing ACT DR6 data and surfacing non-trivial scale- and pair-dependent signal behavior. Together, the two systems span the range from well-defined optimization targets to the messier reality of genuine data analysis. Cosmology, with its mix of clean benchmarks and complex real datasets, turns out to be a useful proving ground for testing how far AI autonomy can actually reach.