"It just feels really dangerous. The product team shows up with a vibe-coded app and they're like, where can we put this? And I was like, you can put it in my production environment, where all the rest of the stuff is. But if something goes wrong with it, it will destroy the company."

So can AI-generated code destroy a company? For a reliability engineer, the honest answer is: it depends entirely on whether your infrastructure was built to absorb it. Most are not.

This perspective comes from a veteran with two decades of experience across planetary-scale infrastructure, the observability sector, and financial reliability. It is a dynamic we have heard more than once on calls like this: AI-enabled non-engineers bypassing traditional dev cycles to clear the backlog, leaving reliability engineering to manage the potential fallout. Hearing it play out, the image was clear — an expert who has seen it all, metaphorically grabbing a lawn chair and a cold beer to watch an unavoidable, self-inflicted meltdown.

Production is where everything lives. When something breaks, the failure cascades through systems while six or seven overlapping monitoring tools generate conflicting signals — metrics split across three systems, infrastructure that is part Terraform and part click-ops, with changes nobody has fully mapped. There was no clean answer. There rarely is when vibe coding and AI-generated code arrive faster than the infrastructure designed to hold them.

And this is exactly what AI-generated code is impacting. Not only the quality of the code, but also the rate at which it arrives at the door of infrastructure that was built for a slower world. Product teams experience AI coding tools as a multiplier on their output. Reliability engineers experience them as a multiplier on their exposure to technical debt. The velocity is shared. The liability is not.

He put it plainly: the "AI stuff" is making everybody a developer. It's moving the bottleneck in software delivery from design-and-build to getting it out and managing the impact later. The constraint used to be who could write the code. Now the constraint is who can safely run what gets written. And because reliability engineering teams aren't growing, that gap is widening.

The answer of course is not to slow development down. That horse has long left the barn. AI coding agents are producing software at a rate and volume no reliability engineer can personally absorb. The answer to agentic coding is, well, agentic reliability. Systems that learn the environment before anything breaks, detect problems before alerts fire, and investigate without requiring someone to have written the answer down in advance.

If you are the person who gets asked "where does this go" more often, from more directions, with less notice than before, that is not a deployment question. It is a signal that the gap between how fast your organization can create software and how well it can operate it is becoming a liability.