“Soar: An Architecture for General Intelligence” was published by John E. Laird, Allen Newell, and Paul S. Rosenbloom in the journal Artificial Intelligence, volume 33, in 1987, building on a system first assembled in 1982. Soar was Newell’s vehicle for one of the boldest aims in the field: a single architecture that could serve as a unified theory of cognition, performing any intelligent task rather than one narrow problem. The source cited here is “Introduction to the Soar Cognitive Architecture,” a current firsthand overview by Laird that describes the same system as it stands today.
Soar treats every task as search through a problem space, a set of states and operators, an idea Newell traced back to the General Problem Solver. Its knowledge lives in production rules, condition-action pairs that fire in parallel when their conditions match what is in working memory. When Soar reaches a point where it cannot decide what to do next, it hits an impasse and automatically sets up a new subgoal to resolve it. A learning mechanism called chunking then summarizes how the impasse was resolved into a new rule, so the same difficulty is handled directly next time.
Soar has been used for decades in cognitive modeling, in AI research, and in large simulations such as training environments with many autonomous agents. Alongside ACT-R it is one of the two most influential cognitive architectures, and it remains an active open-source project maintained at the University of Michigan.
Why business readers should care: Soar is a long-running attempt to build a general, learning, reasoning agent out of explicit rules and goals, the symbolic counterpart to today’s learned models. Its mix of rule-based reasoning, automatic subgoaling, and learning from experience anticipates many ambitions of current AI agents.