SkillGate Boosts Agent Performance by Improving Skill Selection Training

Modern AI agent frameworks increasingly rely on "skills" – procedural knowledge files – which agents access on demand from vast libraries. A critical challenge arises when the agent's policy itself must decide which skill to read during an episode, as existing training signals fail to adequately teach this selection process. The core issue, termed "selector credit starvation," occurs because standard outcome-rewarded reinforcement learning distributes credit too broadly, causing the few tokens responsible for skill selection to receive insufficient or even misleading feedback, especially in long task sequences.This research introduces SkillGate, a novel training approach designed to overcome this limitation. SkillGate partitions the token support into two distinct credit channels: one for execution tokens, receiving outcome credit, and a separate, action-local advantage channel specifically for skill-naming tokens. This local advantage is positive only when the correct skill is chosen for a given trajectory.Experiments on five agentic benchmarks, using a 16-candidate skill slate, demonstrated significant improvements. SkillGate elevated a 9B policy's trial success from 40.8% to 53.2%, outperforming an identical budget spent on outcome reward alone. It also reduced exposure to misleading candidates by two-thirds and led to fewer skill reads, highlighting its efficiency and effectiveness in teaching agents to make better skill selection decisions.