Scientific discovery, especially in biomedical research, is like finding a needle in a haystack. A notable example is large-scale genetic perturbation experiments, a crucial part of drug design, where the goal is to find a small subset of many genes that yield a particular phenotype (e.g., regulating the production of Interleukin-2). Here, we develop BioDesignAgent, an AI agent that can effectively design genetic perturbation experiments. Our AI agent is based on large language models, which has rich prior biological knowledge and can generate explainable rationales while selecting genes to perturb. BioDesignAgent attains an average of 23% improvement in detecting strong phenotypes across five datasets (of which one is unpublished and therefore guaranteed to not appear in the language model’s training data) compared to existing Bayesian optimization baselines. Additionally, it is uniquely able to predict gene combinations to perturb based on single-gene perturbation outcomes, a task so far not explored in this setting. Overall, our approach represents a simple new paradigm in the computational design of biological experiments, aimed at augmenting scientists’ capabilities and accelerating scientific discovery.