If the point of the agent is to perform "reasoning" for operators, being able to understand the "train of thought" is part of the output. Removing access to this (efficient for the model or not) is removing value for the customer. AI shrinkflation.

Comparison of "neural space" and "chain of thought", and why CoT was actually worse.