They are pretty close. LLMs can write the code the solve a sudoku, or leverage an existing solver, and execute the code. Agent frameworks are going to push the boundaries here over the next few years.