Agent skills work for a very specific reason: they turn messy past experience into a clean procedure the agent can follow.

The researchers gave agents the same past trajectories in 2 forms: Workflow Memory, which keeps more execution detail, and a distilled SKILL.md.

The skill version performed 6.06 percentage points better than Workflow Memory.

Because the agent was not getting more experience. It was getting the same experience packaged better.

Their trajectory analysis makes the mechanism clearer: 65.7% of skill cases worked through procedural anchoring, while only 4.5% worked by supplying missing knowledge.

So skills mainly help with execution: what to do first, which tools to use, what to verify, and which mistakes to avoid.

This also explains the failure mode. A skill can still hurt when it is used in the wrong situation or followed too rigidly.

Overall takeaway, self-improving agents need better distillation and application of experience, not just bigger memory libraries.