Paper Questions Adding Procedural Skills to AI Agents

Loading AI agents with endless procedural skills without strict utility filtering creates context clutter and degrades overall reliability.

• –Traditional evaluations focus exclusively on average task success while ignoring latency, prompt costs, and the severity of execution failures.
• –Uncurated skill sets cause selection collisions and increase prompt overhead, often hurting baseline reasoning capabilities.
• –Building robust agents requires selective, dynamic skill loading and domain-specific validation rather than static skill accumulation.