Your RL-trained search agent hits the accuracy target a...

The agent is behaving correctly according to its reward function, which lacks a penalty for search calls. To fix this, introduce a penalty term λn to the reward function. The coefficient λ should be set based on the cost per call versus the value of accuracy. While a hard inference cap can recover the budget immediately, it may truncate the multi-hop questions the agent was designed for. Therefore, the recommended approach is to ship the cap, retrain the agent with the new reward, and compare performance at matched call counts.