Agent frameworks increasingly package procedural knowledge as skills—instruction files an agent reads on demand, with public libraries now hosting thousands. Choosing which skill to read has become a decision the policy itself makes mid-episode, yet no existing training signal directly addresses it.
The researchers identify a structural problem they call selector credit starvation: under a broadcast, sequence-level advantage, the few tokens that name the chosen skill carry a vanishing share of the loss, and the credit they inherit becomes increasingly wrong-signed as trajectories lengthen. A correct choice is punished whenever execution after it fails, even though the choice is among the most valuable decisions in the trajectory.
SkillGate removes this failure by construction. It partitions token support into two disjoint credit channels: outcome credit reaches only execution tokens, while a separate action-local advantage reaches exactly the skill-naming tokens, positive only when the trajectory's single read is the correct one.
On five agentic benchmarks with a 16-candidate slate, SkillGate lifts a 9B policy from 40.8% to 53.2% trial success—well ahead of the identical budget spent on outcome reward alone—while cutting exposure to misleading candidates by two thirds and reading fewer skills.