Researchers have proposed a new method called SkillGate to improve the performance of long-horizon agents in decision-making tasks. These agents use skills, which are pre-written instruction files that can be read on demand, but they struggle with selecting the right skill at the right time. The current approach, using outcome-rewarded reinforcement learning, is not effective due to a problem called selector credit starvation. SkillGate addresses this issue by separating the credit for the agent's decisions into two channels: one for the outcome and another for the action-local advantage. This allows the agent to focus on making correct skill selections without being penalized for subsequent failures. The method has been tested on five benchmarks and shown to improve performance significantly, reducing exposure to misleading candidates and reading fewer skills. --- Why it matters: This matters because long-horizon agents are increasingly used in real-world applications such as robotics and autonomous vehicles, where decision-making under uncertainty is crucial. SkillGate's improvement can lead to better task completion rates and reduced errors in these domains. Source: https://arxiv.org/abs/2608.18852