Vibe coding is fun right up until pager duty at 2:13 a.m. and nobody can explain why production just lit itself on fire. Simon Willison’s warning lands because it’s not anti-AI drama—it’s a blunt diagnosis of where agentic engineering is headed when teams confuse “it works once” with software reliability.
Right now, too many teams are shipping AI code generation output they barely understand, then acting shocked when debugging turns into archaeology. Prompts are not architecture, and vibes are not control flow. If your system behavior depends on probabilistic guesswork stitched together by optimistic engineers, you didn’t build an AI enterprise stack—you built a demo with a hidden fuse.
The upside is obvious for anyone building developer tools: this is the opening for AI testing, traceability, constraint engines, and control-flow-first LLM engineering frameworks. The next big AI software winners won’t be “faster code typing” products; they’ll be the boring, essential infrastructure that makes agentic code auditable, reproducible, and survivable in production. That’s where serious ai consulting firms will make real money too, from ai consulting los angeles boutiques to global integrators.
My hot take: deterministic code isn’t dead, but blind trust in generated code should be. We’re absolutely building a house of cards when we let agents ship without guardrails, and the bill is coming due in outages, security incidents, and maintenance debt. Rating: 9.3/10 story—because this is one of the few AI debates that actually matters before the next catastrophe memo hits Slack.
Stay sharp. — Max Signal
