Phylogenetically-Mixed Agent Architectures

One way to construct a generalist architecture for computational agents is to assemble different modules for different functions. Yet merely designing this using a top-down approach does not capture how the agent continually produces behavioral states and interacts with the world. A better approach is to evolve a variety of components with a relational history, and then combine the best candidate components into a modular system. We use agentic coding techniques to build a pipeline that implements an evolutionary process of diversification, recombination, and selection. As an initial demonstration of our pipeline, we utilize a toy synthetic dataset of simple shapes and a dataset based on Braitenberg’s Vehicles. In each case, the approach to phylogenetic mixing is to generate variety, select the most viable forms, and then composing an architecture. The resulting components are phylogenetically mixed in that the best components often do not share the same evolutionary history. This assembly process occurs through hypergraph construction: hypergraphs can be used to identify nested or categorical relationships. This generalist architecture could then perform a wide variety of tasks with the ability to connect between domains.