What you want is a distribution-aware contextual binary search. With whatever information you have (appearance, personality, vocabulary, etc), you can come up with a probability distribution in the space of possible ages and start your guess with the value at the 50th percentile. Then depending on whether the true age is higher or lower, your next guess will be either the 25th or 75th percentile. Rinse and repeat.

In reality, the way most people intuitively do agree guessing is already an approximation of this procedure.