Oncomorphic neural agent populations for resource-limited sequential learning
Distributed artificial intelligence (AI) often operates under sequential task exposure, uneven compute, and decentralized coordination. Here, we present a cancer-inspired, or oncomorphic, multi-agent framework in which simulated neural agents can replicate, mutate their neural network architecture, migrate across task environments, undergo ecological turnover, and recruit learning/ecological resources from a finite shared reserve. We evaluate the framework in controlled synthetic nonlinear classification environments in which each agent trains only on its local task, allowing population ecology rather than centralized optimization to determine which neural network architectures persist. For various initial conditions, we find that stronger selection increased the endpoint local accuracy of surviving agent populations. Architecture mutation played a state-dependent role: diverse initial populations performed best at low mutation, whereas clonal large-architecture populations benefited from mutation-generated variation. Selection also increased end-of-run multi-task competence, measured by evaluating surviving agents on all environments without additional training. Recruitment and elevated baseline replication reshaped demographic support while prediction quality remained within a narrow band, consistent with redistribution of finite learning resources. Time-resolved entropy and dominance analyses revealed concentration toward successful architectures, while finite training cycles kept agents in a non-asymptotic learning regime. These results provide proof-of-concept mechanistic evidence that oncomorphic population dynamics may offer a route to decentralized adaptation in engineering applications under bounded local resources.