Agent Based Learning Systems builds simulations where neural-network agents self-organize, specialize, and coordinate, a scientific instrument for understanding AI, institutions, and decentralized systems before they are deployed.