Correct: automatically create a subgoal, in a new problem space, whose job is to decide between the two operators.
Why not the others: a random pick (option 1) throws away information the architecture actually has access to — the subgoal search might discover, say, that one operator leads somewhere the other doesn’t, information invisible if you’d just flipped a coin and moved on. Halting (option 2) treats getting stuck as a bug; SOAR treats it as the normal, expected shape of a hard decision. Applying both (option 4) isn’t coherent for actions with real, possibly conflicting effects on a single state.
The elegance is architectural, not just procedural. SOAR could have bolted on a special-purpose “tie-breaking module” separate from its normal reasoning. It didn’t. An impasse just becomes a new problem: a new state (the tied situation itself), new operators available for reasoning about it (maybe “look up which operator worked last time in a similar state,” itself possibly requiring another subgoal), and a goal test (“an operator preference has been decided”). Resolving that subproblem returns a result to the level above, the original impasse clears, and normal operation resumes — exactly the way a stack frame returns a value and execution resumes in the caller. Deeply nested impasses (a subgoal that itself hits an impasse, recursively) are not a special case to handle; they’re just what the same one mechanism does when applied to itself.
Why “universal.” The name isn’t marketing — Newell’s argument was that every known impasse-type in intelligent behavior (operator tie, operator not applicable, operator’s result unknown, goal not yet reached) reduces to “spin up a problem space for this” with no exceptions required. One mechanism, applied recursively, was Newell’s specific bet about what “unified theory of cognition” should mean architecturally: not many special modules, but one generative mechanism producing all the special-looking behavior as instances.
In a modern LLM agent harness, this rhymes with an agent that, upon finding no clear next tool call (ambiguous instructions, two plausible interpretations), spins up a sub-task — “first determine which interpretation is intended” — using the same reasoning loop as the main task, rather than a bespoke disambiguation routine wired in separately. The parallel is the architectural choice, not a specific implementation: reuse the general reasoning loop for meta-reasoning about the loop itself, instead of hand-coding a separate escape hatch.
Where this goes: next lesson asks what SOAR does with the result of one of these subgoal searches once it’s solved — and it isn’t just “apply it and forget it.”