🛑

Andrey Karpathy (ex-Tesla AI, ex-OpenAI) dropped another truth bomb in a new post. The gist is simple: programming has changed beyond recognition. And it didn't happen gradually, but in a sharp leap literally over the last couple of months.

In his opinion, until December, AI agents for writing code essentially didn't work. But now they do. Models have broken through the quality ceiling, learned to maintain long-term context, show persistence, and independently navigate complex, large-scale tasks.

👉🏻 Andrey's practical case:
Over the weekend, he needed a local dashboard for video analytics from home cameras. He didn't open an editor; instead, he wrote a single prompt in English to an agent:
Here is the IP, login, and password for my local server. Log in, set up SSH keys, install vLLM, download and run the Qwen3-VL benchmark. Set up a server endpoint for video inference, throw together a basic Web UI, test everything, wrap the services in systemd, save notes for yourself from memory, and write me a Markdown report.


The agent went off to work for 30 minutes. During that time, it encountered a bunch of bugs on its own, googled solutions, wrote code, debugged, set up services, and returned with a fully ready result. Karpathy didn't even touch the code.
What three months ago was a weekend pet project is now a background task that takes half an hour—you launch it and go grab coffee.

🤖 Agentic Engineering
According to Karpathy, we are right now saying goodbye to the concept of "a person typing instructions in a text editor."

Now development is when you spin up instances of AI agents, give them tasks in natural language, and simultaneously manage and review their work. The winner is the one who learns to climb higher up the layers of abstraction and manage long-lived orchestrators that control a crowd of parallel AI coders, providing them with the necessary tools and context.

⚖️ Where does the human fit in?
This is still not the magic of "make me a money button." Agents still critically need:
🟢 High-level direction and architectural vision.
🟢 Engineering taste and judgment (that same judgement).
🟢 Iterative prompting.

This works best where the task has a rigid specification and functionality can be easily verified with tests. The main skill now is intuition in decomposition. You must be able to break a complex system into atomic pieces that can be safely delegated to an agent, and control their interfaces.

Essentially, syntax is no longer worth almost anything. Architecture, decomposition, and the ability to manage complex abstractions are worth everything.