SOAR

SOAR (State, Operator, Result) is a cognitive architecture from Allen Newell that models problem-solving as cycles: you're in a state → apply an operator → land in a new state, repeated toward a goal. When the agent can't pick an operator, it hits an impasse and creates a subgoal to resolve it; successful action sequences are learned as new rules through a mechanism called chunking. SOAR predates modern AI by decades, but its principles — hierarchical decomposition and learning from successful sequences — still inform how long-running agents are designed.