Abstract
Online behavioral experiments rely on the assumption that participants are human. Advances in large language models(LLMs) challenge this assumption, but previous demonstrations have largely focused on surveys or agents built for individualexperimental tasks. Here, we introduce the Artificial General Experiment Agent (AGEA), a general-purpose agent developedentirely through natural language prompting using a consumer-level LLM subscription. Without custom training for anyspecific task, AGEA completed twenty-six online behavioral experiments across seven hosting platforms, independentlyinferring each task’s structure. Benchmark comparisons across three cognitive domains, using the N-back, Tower of London,and Wisconsin Card Sorting Test, showed performance within typical human ranges, including changes in accuracy andresponse time with experimental conditions. It further completed, with human-like behavior, seven held-out studies preparedin the spirit of adversarial collaboration between the co-authors. These findings show that agent participation can extendacross experimental paradigms and platforms. As general-purpose agents become cheaper and more capable, they pose agrowing risk to the integrity of online behavioral research, and safeguards must keep pace.