Evaluating Skills, Not Just Agents - Skill Lift

Most review processes for agent skills still treat them as documents. They check structure, style, and security. They do not answer the deployment Liftquestion: does this skill actually help a live agent complete real tasks under the same model, sandbox, and grading policy?

ACES (Agentic Continuous Evaluation of Skills) answers that question with paired live trials. For each task the system runs the agent once with the target skill and once without it. Trajectories are normalized, graded on multiple runtime metrics, and compared. The central measure is Skill Lift: the measured improvement the skill produces under fixed conditions.

The results show that static scans and live evaluation measure different things. Structural and style checks catch authoring issues. They do not observe discovery failures, incorrect invocation, workflow drift, or inefficient tool use. Those process failures only appear when the skill is exercised by a live agent. Across real production skills, mean Skill Lift is positive and the largest gains appear precisely in the metrics that document scans cannot see.

The practical implication is that skills should be treated as executable artifacts with continuous evaluation, not as static packages that pass a one-time review.

This Ledger Entry expands how the reader thinks by shifting skill governance from authoring quality to demonstrated runtime contribution.