A framework for a new supervised learning algorithm, called Insight that represents a situated agent, is sketched. The envisioned agent should be able to predict the behaviour of its finite, deterministic, fully observable, singleagent environment represented by a Finite State Machine. By generalising the observed state transitions, which are only a fraction of all possible ones, the agent should become capable to predict the full state transition table. States are not – as it is common in automata theory – simply enumerated but are represented as structured entities, as vectors of primitive values, called state aspects. State vectors are mapped one-to-one to the agent’s perception apparatus, each state aspect corresponding to a single sensor. The envisioned agent should describe observed transitions in terms of a self-invented (concept) vocabulary of objects (nouns), their properties and interrelations (adjectives) and changes (verbs).