Agent-based simulations are rule-based models traditionally used for the simulations of complex systems. In this paper, an algorithm based on the concept of agent-based simulations is developed to predict the lowest energy packing of a set of identical rigid molecules. The agents are identified with rigid portions of the system under investigation, and they evolve following a set of rules designed to drive the system toward the lowest energy minimum. The algorithm is compared with a conventional Metropolis Monte Carlo algorithm, and it is applied on a large set of representative models of molecules. For all the systems studied, the agent-based method consistently finds a significantly lower energy minima than the Monte Carlo algorithm because the system evolution includes elements of adaptation (new configurations induce new types of moves) and learning (past successful choices are repeated).