The outcomes of installs and upgrades may be totally different every time, even with the very same mannequin, but it surely will get loads worse if you happen to improve or swap fashions. When you’re supporting infrastructure for 5, 10, or 20 years, you will be upgrading fashions. It’s exhausting to even think about what the world of generative AI will appear like in 10 years, however I’m positive Gemini 3 and Claude Opus 4.5 won’t be round then.
The hazards of AI brokers enhance with complexity
Enterprise “purposes” are not single servers. In the present day they’re constellations of techniques—internet entrance ends, software tiers, databases, caches, message brokers, and extra—typically deployed in a number of copies throughout a number of deployment fashions. Even with solely a handful of service varieties and three primary footprints (packages on a conventional server, picture‑primarily based hosts, and containers), the combos develop into dozens of permutations earlier than anybody has written a line of enterprise logic. That complexity makes it much more tempting to ask an agent to “simply deal with it”—and much more harmful when it does.
In cloud‑native retailers, Kubernetes solely amplifies this sample. A “easy” software may span a number of namespaces, deployments, stateful units, ingress controllers, operators, and exterior managed companies, all stitched collectively via YAML and Customized Useful resource Definitions (CRDs). The one sane approach to run that at scale is to deal with the cluster as a declarative system: GitOps, immutable photos, and YAML saved someplace exterior the cluster, and model managed. In that world, the job of an agentic AI is to not scorching‑patch operating pods, nor the Kubernetes YAML; it’s to assist people design and check the manifests, Helm charts, and pipelines that are saved in Git.