Skill drift changes production agents without a redeploy on Gemini Enterprise

On 2026-11-02, the AI section of DEV Community published an article titled, "No redeploy, but your agent changed: Skill Drift on Google Cloud". It describes a situation that, in the author's account, is easy to create on Gemini Enterprise Agent Platform today. First, you build a Managed Agent and attach a skill from Skill Registry called invoice-triage. It works well and you ship it. Two weeks later, someone on another team updates invoice-triage to handle a new vendor format, tests it with their own agent, and it works for them. That same afternoon, the production agent begins misrouting invoices, even though nobody touched it and no deploy appears in the logs. The author calls this phenomenon skill drift. The article then proposes a lockfile pattern, based on the public Managed Agents API and Skill Registry documentation, for pinning skill versions and measuring the context overhead that mounting many skills introduces. The available material includes only this one report, so there is no corroborating coverage yet and no independent explanation of what causes skill drift.