Nanoparticles have the potential to modulate both the pharmacokinetic and pharmacodynamic profiles of drugs, thereby enhancing their therapeutic effect. The versatility of nanoparticles allows for a wide range of customization possibilities. However, it also leads to a rich design space which is difficult to investigate and optimize. An additional problem emerges when they are applied to cancer treatment. A heterogeneous and highly adaptable tumour can quickly become resistant to primary therapy, making it inefficient. To automate the design of potential therapies for such complex cases, we propose a computational model for fast, novelty based machine learning exploration of the nanoparticle design space. In this paper, we present an evolvable, open-ended agent-based model, where the exploration of an initially small portion of the given state space can be expanded by an ongoing generation of adaptive novelties, whenever the simulated tumour makes an adaptive leap. We demonstrate that the nano-agents can continuously reshape themselves and create a heterogeneous population of specialized groups of individuals optimized for tracking and killing different phenotypes of cancer cells. In the conclusion, we outline further development steps so this model could be used in real-world research and clinical practice.