Leo's blog

For the last two years, the AI conversation has been dominated by “autonomy.” We’ve been sold a vision of the “autonomous agent”—the digital employee that you give a goal to, and it just handles it.

It’s a compelling story. It’s also, for most businesses, complete theatre.

The reality of deploying AI in a real-world production environment is not a problem of “better LLMs” or “smarter prompts.” It is a problem of governance. When you move from a demo to a deployment, you realize that autonomy without a control plane is just a fast way to corrupt your data and alienate your customers.

The real value is no longer in the model. The value is in the Control Layer.

The Death of Vibe Coding

We are seeing the transition from “Vibe Coding” to “Software Factory Supervision.”

Vibe coding was the honeymoon phase: writing a few prompts, seeing a plausible result, and assuming the system “understood” the intent. But vibes don’t scale. Vibes don’t pass an audit. Vibes don’t handle edge cases in a legacy database.

The IDE is evolving. It is no longer just a place to write code; it is becoming a control room. The shift is from writing the logic to supervising the factory that generates the logic. The operator’s job is no longer to be the best coder in the room, but to be the best architect of the constraints.

The Autonomy Trap

The market is currently flooded with “AI-operated companies.” In 99% of cases, this is fluff.

The real market—the one that actually moves the needle—is governed AI-operated functions.

The goal isn’t to replace the human decision-maker; it’s to automate the 90% of the grunt work leading up to that decision, while ensuring that the human still owns the liability and the final “Yes/No.” If you remove the human from the loop entirely, you aren’t building a business; you’re building a lottery ticket.

The New Stack: From Prompt to Plane

If you want to get actual value from AI today, you stop obsessing over the prompt and start obsessing over the infrastructure:

1. Data Hygiene as a Prerequisite: AI doesn’t fix messy data; it accelerates the speed at which messy data creates wrong decisions. The “AI Strategy” must start with a “Data Cleanup Strategy.”
2. Role Separation: You cannot have the same agent building the solution and verifying the solution. Trust boundaries must be hard-coded. Builder $eq$ Verifier.
3. Local Runtime Stability: The cloud is for training; the edge (or local runtime) is for execution. Privacy, latency, and reliability demand a move back toward local-first operator-grade hardening.

The Bottom Line

The “magic” of AI is wearing off, and that is exactly where the real work begins.

The winners won’t be the ones who found the best prompt. They will be the ones who built the most robust control plane. They will be the ones who realized that the model is just a commodity—and that the governance of that model is the actual product.