A longitudinal empirical study using primary-source version-history data challenges the assumption that agent skill and tool maintenance can be automated end-to-end. The researchers tracked who performs maintenance work on agent skills over time. Human contributors drive the majority of meaningful updates, including retiring outdated tools, correcting input-and-output format definitions, and adjusting tool behavior. Automated tooling helps at the margins, but governing skill quality remains a human responsibility. For teams building tool ecosystems or self-evolving agent pipelines, this has direct implications. Skill rot, where a registered tool drifts out of sync with the systems it wraps, is a failure mode that automated checks do not catch in full. The version history shows that humans notice and fix subtle breakdowns, such as a changed service interface or a shift in downstream data shape, that automated test suites miss. Tests miss these because they were written against stale assumptions. Skill maintenance needs an owner, a schedule, and a visible queue, the same infrastructure you would give a software team managing a live service. Treating skill registries as write-once artifacts and assuming the agent or a monitoring layer will self-correct is a governance gap. That gap shows up as silent degradation in agent output quality.