Thinking-in-Bets is a training-free method that prevents agents from abandoning correct work when challenged by a user. By maintaining a private record of claims and confidence before reading feedback, agents can distinguish between real mistakes and user error.
HOW THIS AFFECTS YOU
●
builderYou can reduce agent sycophancy in your products by implementing a private verification step before feedback processing.
●
researcherThis presents a way to improve agent robustness without additional fine-tuning.