Abstract: A system for agent training may receive one or more tasks for a character to complete in one or more training sessions of an environment, where the character is controlled by a player that is selectable between a human user and an agent implementing an active model. The system may dynamically switch the player between the human user and the agent during the training sessions, where at least one condition for switching the player is based on an introspection metric associated with at least one of a confidence of the agent or observed actions when the player is the human user. The system may further capture observations from the character in the environment and corresponding actions of the character in the environment from the player during the training sessions as experience data, and train one or more models for the agent based on the experience data.