Google put a cost curve on the thing I have been betting Falkster on. A whitepaper, The New SDLC With Vibe Coding, co-authored by Addy Osmani, a Director at Google Cloud AI, draws a hard line between vibe coding and agentic engineering. It is not subtle about which one survives contact with a real codebase. The slogan from the paper is the whole thesis in four words: structure scales, vibes don't.
The chart is the argument
Skip the prose and look at the curve. Cumulative cost of ownership on one axis, features shipped over time on the other. Two lines. Vibe coding starts at almost zero, because there is no setup, no specs, just output. So early on it wins loudly, because the only metric it optimizes is speed to first working thing. Agentic engineering starts higher, because you paid for structure before you saw anything work.
Then the lines cross. The vibe coding curve bends upward and keeps climbing, because every shortcut compounds into rework, regressions, and code nobody can safely change. The agentic curve flattens, because the structure you bought early keeps catching the problems the vibe loop would have shipped. Google's number for the gap past the crossover is 3 to 10x more per feature.
It is a CapEx versus OpEx decision
The cleanest way to read the chart is as an accounting choice. Vibe coding is low CapEx, high OpEx: you pay nothing to start and everything to maintain. Agentic engineering inverts it, a real upfront cost to build the substrate, then low marginal cost per feature because the structure does the carrying. The crossover point is just the moment the rising OpEx of vibes overtakes the amortized CapEx of structure.
The trap is the quiet graduation
Nobody decides to take the high-OpEx path. They drift into it. A prototype gets a demo, the demo gets a customer, the customer gets a second feature, and at no point did anyone stop to say this is a product now, it needs structure. The thing graduates from prototype to production without ever paying the structure tax, and you hit the crossover without noticing until maintenance is eating the team alive. Cheap to build is not the same as cheap to own.
What agentic engineering actually means in practice: write the eval as the spec so the definition of done is a test the agent has to pass, treat prompts as versioned operations, and keep vibe coding for internal tools and throwaways where the crossover never arrives.
This week, take the last thing your team vibe coded into production and ask one question: have we paid the structure tax on it, or are we just not at the crossover yet? If it has no eval, no guardrails, and no review, you are not saving money. You are accruing it as debt at 3 to 10x interest. Put structure under the one piece that matters most before the curve turns.