AI News

Tool-using agents are often updated with new interaction data, but their previously estimated action credits can become stale after each policy change. Recomputing credit from scratch requires additional tool calls and environment interactions. This study argues that a change in action value matters only when it can change the ranking of candidate actions. It introduces pairwise branch sensitivity to measure how a policy update affects downstream regions that distinguish two actions, along with a first-order anchored credit-transport estimator that reuses old interventional trajectories. The proposed Decision-Sufficient Credit Gate, or DSC-Gate, chooses among reusing historical credit, transporting it, and resampling it. In experiments, branch sensitivity explained credit drift better than global policy distance, and credit transport reduced estimation error when enough historical data was available. Its decision-making benefit was concentrated on updates affecting action-distinguishing branches. On an independent test set, DSC-Gate changed mean regret by only +0.00004 relative to a gap-based gate while reducing mean new tool steps from 472 to 286, a 39.4% reduction. The same pattern appeared after a real tool-agent parameter update, suggesting that agents may not need to recompute action credit after every policy update.