A second paper calls out rotting AI skill banks

Agents rarely check whether their saved skills still work, SkillForge argues. This matters for agents meant to run long enough to build up a skill bank. Skills that used to work start blending in with ones that don't, and nothing flags the difference.

The paper points to SkillRL, an existing method for extracting skills during reinforcement learning. Its skill bank only grows because nothing verifies a stored skill against the current environment.

SkillForge's fix has three pieces. It makes skill usage explicit during interaction. That lets the training process optimize both the actions an agent takes and its decisions about which skill to invoke. It adds evidence-based verification, checking whether a skill still holds up. It also draws on multiple pathways to induce skills, rather than one fixed extraction method. Tested against SkillRL on ALFWorld, WebShop, and AppWorld, SkillForge wins consistently across all three.

This is the second paper in as many weeks calling out silent skill staleness in agent systems. One paper flagging the issue is a data point. Two papers raising it within weeks of each other reads like the field starting to treat skill-bank hygiene as a requirement, not a nice-to-have.

The open question is whether agent frameworks start shipping skill verification as a built-in step. The alternative is that builders only notice when an agent pulls from a bank of dead skills, with no error explaining why.