"LLMs can't do _" Think of agent harnesses as guided brute forcing tools. If we can define a problem with good boundaries and an evaluation function, shove it in an inference in a loop, then we might get better-than-fuzzing traversal to novel results.