AutoTutor is a natural language tutoring system that has produced learning gains across multiple domains (e.g., computer literacy, physics, critical thinking). In this paper, we review the development, key research findings, and systems that have evolved from AutoTutor. First, the rationale for developing AutoTutor is outlined and the advantages of natural language tutoring are presented. Next, we review three central themes in AutoTutor's development: human-inspired tutoring strategies, pedagogical agents, and technologies that support natural-language tutoring. Research on early versions of AutoTutor documented the impact on deep learning by co-constructed explanations, feedback, conversational scaffolding, and subject matter content. Systems that evolved from AutoTutor added additional components that have been evaluated with respect to learning and motivation. The latter findings include the effectiveness of deep reasoning questions for tutoring multiple domains, of adapting to the affect of low-knowledge learners, of content over surface features such as voices and persona of animated agents, and of alternative tutoring strategies such as collaborative lecturing and vicarious tutoring demonstrations. The paper also considers advances in pedagogical agent roles (such as trialogs) and in tutoring technologies, such semantic processing and tutoring delivery platforms. This paper summarizes and integrates significant findings produced by studies using AutoTutor and related systems.