Claude slop article, and also (speaking as a 20 year IaC vet) the premise is misplaced.
Declarative infra with drift observability has been readily available, it's just that cloud devs -- to Claude's point -- have largely written and used Chef, Ansible, TF, etc. imperatively.
Several of the examples (such as S3) are true only when “doing it wrong”. But, it's not exactly wrong, it's more like imperative programming vs. functional programming. It's a choice, with tradeoffs.
Claiming one needs LLMs to read metadata from the CSP is wonky as well, unless you're shopping an LLM thing to read metadata from CSPs.
We've always likened correct use of CSP to using a holodeck on Star Trek. If something in the holodeck's world is wrong, you don't try to fix it in (virtual) world, you try to fix it in the holodeck program so the correction manifests in the (virtual) world. That is the power and, really, entire point of a CSP. If you're not using — and getting value from — that power, you're burning money compared to your own datacenter.
Just out of curiosity, are you familiar with infrastructures.org? It's a manifesto on how to bootstrap infrastructure from a team that would go on to invent IaC tooling. I read the site thoroughly a few years back and wondered how much differently you'd order the bootstrapping process for modern infrastructure.
I'll admit up front that my infra experience is just limited to my own projects, but even I have used terraform, chef, puppet and ansible enough to know that this article's entire premise is flawed from the start. Did some marketing dweeb write the prompt?
TF's first step on every run is to observe the current state of your infrastructure, you can't miss it!
And since when are we calling every piece of automation an agent?
Declarative infra with drift observability has been readily available, it's just that cloud devs -- to Claude's point -- have largely written and used Chef, Ansible, TF, etc. imperatively.
Several of the examples (such as S3) are true only when “doing it wrong”. But, it's not exactly wrong, it's more like imperative programming vs. functional programming. It's a choice, with tradeoffs.
Claiming one needs LLMs to read metadata from the CSP is wonky as well, unless you're shopping an LLM thing to read metadata from CSPs.
We've always likened correct use of CSP to using a holodeck on Star Trek. If something in the holodeck's world is wrong, you don't try to fix it in (virtual) world, you try to fix it in the holodeck program so the correction manifests in the (virtual) world. That is the power and, really, entire point of a CSP. If you're not using — and getting value from — that power, you're burning money compared to your own datacenter.
That part of the pitch, I agree with!