A few months ago, Andrej Karpathy admitted he had never felt more behind as a programmer. Coming from someone who helped build modern AI, that statement grabs your attention.
He noticed that AI coding tools suddenly became good enough that he rarely needed to correct them. He started “vibe coding”—trusting the AI to write the code while he directed the overall project. This realization marks a major shift in how we build technology. We are entering the era of Software 3.0 and agentic AI.
Here is what that means for teams, managers, and the future of software.
The Three Eras of Computing
To understand this shift, look at how computing has evolved:
- Software 1.0: Programmers write explicit rules and code using languages like C++ or Python.
- Software 2.0: Programmers curate datasets and train neural networks to learn the rules.
- Software 3.0: You use natural language to prompt a large language model (LLM). The LLM acts as the computer, interpreting your context and doing the heavy lifting directly.
In Software 3.0, traditional apps often become unnecessary. Karpathy shared an example of an app he built called MenuGen. It took a photo of a text-only restaurant menu and used multiple cloud services to generate images of the food. It was a complex, multi-step setup.
But in the Software 3.0 world, you skip the app entirely. You just give the photo to an AI model and ask it to overlay images onto the text. The neural network does the work directly. In the future, neural networks will likely become the main processor, and traditional computer chips will just assist them.
Vibe Coding vs. Agentic Engineering
If AI can do the work, what do software engineers actually do? Karpathy breaks it down into two concepts.
First is vibe coding. This raises the floor. Anyone can use AI to build basic software easily.
Second is agentic engineering. This maintains the quality bar. Professional software still needs to be secure, reliable, and functional. Agentic engineering is the discipline of managing slightly unpredictable AI agents to build high-quality systems much faster than before.
You manage agents almost like a team of eager interns. They handle the details, like remembering specific code syntax or API rules. But you design the structure, provide oversight, and enforce the rules.
Jagged Intelligence and What Automates Next
AI models are incredibly smart in some areas and surprisingly clueless in others. Karpathy calls this “jagged intelligence.” An AI might find complex security flaws in a massive codebase, but if you ask it how to get to a car wash 50 meters away, it might tell you to walk instead of drive.
This happens because AI models get exceptionally good at verifiable tasks. Labs train them heavily on things with clear right or wrong answers, like math and coding. If your business problem is verifiable, AI can likely automate it right now. But because of these blind spots, you still need to stay in the loop to guide the tools.
Moving to Agent-Native Systems
Right now, most of our digital world is built for humans. We read documentation, click through settings, and configure services manually.
As agents do more work, we need agent-native infrastructure. Instead of documentation telling a human how to install software step-by-step, it should just provide the exact text you paste to your agent so it can do the installation. We are moving toward a world where your agent talks to other agents to coordinate meetings, deploy code, and handle digital tasks.
The Human Role: Taste and Understanding
With agents taking over the execution, what human skills matter most? Taste, judgment, and understanding.
As Karpathy noted, “You can outsource your thinking, but you can’t outsource your understanding.” You still need to know what you want to build, why it matters, and how the underlying systems work. AI cannot replace human understanding. It simply gives you better tools to turn that understanding into reality.