Create N agents with random genomes containing hyperparameters, optimizers, and data strategies.

Train models using each agent's genome configuration. Parallel execution on distributed GPUs.

Evaluate models on accuracy, efficiency, and stability metrics to compute fitness scores.

Select top-performing agents based on fitness. Typically 20-30% survive to reproduce.

Breed surviving agents using uniform or single-point crossover to create offspring genomes.

Apply random mutations to offspring to introduce genetic diversity and exploration.

New population becomes the current generation. Cycle repeats for continuous evolution.