Researchers have introduced Mechanist, an agentic system designed to automate the discovery of the underlying mechanisms of artificial intelligence. To facilitate autonomous scientific inquiry, the authors constructed a specialized scientific knowledge graph comprising 13,000 AI mechanism studies and integrated a multidisciplinary database containing 43 million papers across 26 fields. The system utilizes a curated library of 32 foundational analysis methods to ensure reliable experimental execution. Mechanist outperformed Claude Code and previous AI-scientist frameworks in hypothesis quality and experimental reliability. Through four case studies, the system demonstrated significant capabilities: it identified a counterintuitive safety risk where unsafe traits transfer to fine-tuned models via seemingly benign data across modalities; it formulated a comprehensive mechanism theory of belief to explain how models represent knowledge and infer the beliefs of others; and it applied these theories to develop targeted interventions that enhance model performance. Furthermore, Mechanist proved capable of advancing interdisciplinary discovery through mechanistic design, offering a more efficient alternative to traditional, computationally expensive generate-and-rerank paradigms. By bridging the gap between rapid AI development and manual interpretability research, Mechanist serves as a scientific instrument for understanding complex model behaviors and optimizing AI design.