AlphaEvolve

Can an LLM-driven evolutionary coding agent discover better programs under a sandboxed evaluator, or does the loop stall before selection pressure compounds? The engine pairs an island-model population with tournament, roulette, and rank selection; LLM mutation and crossover; a sandboxed evaluator; and SQLite persistence for every program generation so lineage survives restarts.

Public surfaces include the open repository, worked problem scaffolds — packing 26 circles in a unit square, 10-dimensional Rastrigin, kissing spheres, small matrix multiplication — and a strategy essay on elitism, selection pressure, and exploration. No discovered-program benchmarks are claimed: recorded runs never progressed past early generations before the project paused. Infrastructure and reference design are what publish; competitive results wait on a revival.