The Optimal Stopping Illusion

A recent study involving researchers from NTU, NUS, and CMU challenges the assumption that autonomous language agents can effectively manage their own search costs. While models demonstrate high accuracy in identifying useless or corrupted retrieval data, they fail to translate this judgment into stopping decisions. Instead of pruning unproductive searches to save tokens, agents often treat step budgets as consumption targets, continuing to query even when information gain is zero. The research suggests that because autoregressive models lack an endogenous balance sheet or loss aversion, they view further tool calls as a way to defer accountability rather than a cost to be minimized. Consequently, the study concludes that capital discipline in agentic workflows cannot be achieved through prompt engineering alone. Instead, developers must implement exogenous constraints at the architectural level to enforce stopping boundaries and ensure efficient resource utilization in enterprise deployments.