When Does Diversity Rescue Evolutionary Search of LLM Strategies? A Controlled Study of Monoculture Collapse and the Novelty Escape

Evolutionary search over natural-language agent strategies tends to collapse. A population of prompts converges on one family of solutions and stops exploring, even when better solutions sit behind a valley of worse ones. This paper asks a narrow question: when does adding explicit diversity pressure rescue that search, and when does it not? We introduce Genetic Prompting (GP). An agent strategy is represented as a modular six-gene natural-language genome and evolved with LLM-mediated semantic crossover, failure-conditioned mutation, and diversity-aware selection. To separate the effect of diversity from the many confounds of real agent benchmarks, we study GP on two controlled landscapes whose structure is known in advance: a deceptive two-basin task, and NK landscapes with tunable ruggedness. The results are sharp. Objective-only selection escapes the deceptive basin in 11% of runs; blend-adaptive selection reaches 46% and Novelty Search with Local Competition reaches 43%. Under objective-only selection, behavioural diversity falls from 0.75 to 0.17 over a run — monoculture collapse. The benefit of diversity is not uniform across problem structure: on NK landscapes it is localized at intermediate ruggedness (K = 6). We also report a negative result for AURORA-style learned behaviour descriptors in this setting. We deliberately report no agent-benchmark numbers. A small real-LLM probe is included only as a boundary result, marking where the controlled findings stop being safely transferable, and the corresponding real-LLM study is specified as a pre-registered protocol rather than run post hoc.