BioDiscoveryAgent is an AI agent designed to create genetic perturbation experiments, leveraging large language models (LLMs) to design new experiments, reason about outcomes, and efficiently navigate the hypothesis space. It outperforms existing Bayesian optimization baselines in predicting genetic perturbations, achieving a 21% improvement across six datasets and a 46% improvement in predicting non-essential gene perturbations. The agent uses an LLM paired with tools for searching biomedical literature, executing code, and prompting another agent to evaluate predictions. It is interpretable at every stage, offering an accessible new paradigm in computational biological experiment design. The agent is tested on two experimental settings: single-gene and two-gene perturbations. In single-gene perturbation, the goal is to identify genes that, when perturbed, result in a specific phenotype. In two-gene perturbation, the agent identifies gene pairs whose combined perturbation leads to the desired phenotype. BioDiscoveryAgent outperforms baselines in both settings, with significant improvements in predicting non-essential genes and in combinatorial gene perturbations. The agent uses a prompt-based approach, incorporating results from previous experiments to guide future rounds. It can access tools for literature search, gene search based on biological databases, and an AI critic to refine predictions. The agent's performance is evaluated across six datasets, with results showing consistent improvements in hit ratios and gene prediction accuracy. BioDiscoveryAgent demonstrates superior performance in both single and two-gene perturbation tasks, outperforming baselines by significant margins. It is interpretable, with reasoning and citations provided for predictions, and can be enhanced through the use of tools. The agent's ability to integrate prior knowledge and observations leads to more accurate and consistent gene predictions. Overall, BioDiscoveryAgent represents a new approach to biological experiment design, offering an accessible and interpretable method for scientists to enhance their experimental efficiency.BioDiscoveryAgent is an AI agent designed to create genetic perturbation experiments, leveraging large language models (LLMs) to design new experiments, reason about outcomes, and efficiently navigate the hypothesis space. It outperforms existing Bayesian optimization baselines in predicting genetic perturbations, achieving a 21% improvement across six datasets and a 46% improvement in predicting non-essential gene perturbations. The agent uses an LLM paired with tools for searching biomedical literature, executing code, and prompting another agent to evaluate predictions. It is interpretable at every stage, offering an accessible new paradigm in computational biological experiment design. The agent is tested on two experimental settings: single-gene and two-gene perturbations. In single-gene perturbation, the goal is to identify genes that, when perturbed, result in a specific phenotype. In two-gene perturbation, the agent identifies gene pairs whose combined perturbation leads to the desired phenotype. BioDiscoveryAgent outperforms baselines in both settings, with significant improvements in predicting non-essential genes and in combinatorial gene perturbations. The agent uses a prompt-based approach, incorporating results from previous experiments to guide future rounds. It can access tools for literature search, gene search based on biological databases, and an AI critic to refine predictions. The agent's performance is evaluated across six datasets, with results showing consistent improvements in hit ratios and gene prediction accuracy. BioDiscoveryAgent demonstrates superior performance in both single and two-gene perturbation tasks, outperforming baselines by significant margins. It is interpretable, with reasoning and citations provided for predictions, and can be enhanced through the use of tools. The agent's ability to integrate prior knowledge and observations leads to more accurate and consistent gene predictions. Overall, BioDiscoveryAgent represents a new approach to biological experiment design, offering an accessible and interpretable method for scientists to enhance their experimental efficiency.