The Evolutionary Layers of Agentic AI

Research By: Vivek - ML & Systems Researcher

In almost every computer science class right now, the conversation has completely changed. Six months ago, students were asking, 'What is AI?' Today, they are asking, 'How do we make it do our assignments?' But you can't just plug a language model into your codebase and expect it to build the architecture. Our research group realized that true automation doesn't happen overnight; it happens in distinct evolutionary stages. To build a high-paying career, we must understand that the evolution of AI isn't actually about intelligence. It is entirely about consequences.

The difference between a tool and an agent.

When you use a basic language model, you are dealing with a Level 1 system. It is brilliant at brainstorming, but it has zero memory of your specific project. To move up to Level 2, we give the AI a library—like a Vector Database—so it can read our notes before it speaks. But it still can't take action. The real jump happens at Level 3, where we give the AI a mouse and keyboard via API access. It can draft a script, but a human must still click 'Run'. The holy grail for engineers is Level 4 and 5, where the AI executes the action completely on its own.

Who takes the blame when it all goes wrong?

The true evolution of AI isn't about how well a model can write code. It is about who takes responsibility when things go wrong. We are shifting from AI that just generates text to AI that actually takes responsibility for its actions. As student architects, we have to learn to code accountability into our projects.

The loop that gets you hired.

At the highest levels, we use what is called the ReAct framework—Reason plus Act. The AI doesn't just output text; it follows a high-speed logical loop. It thinks about the problem, takes an action like checking a database, observes the result, and then revises its thought process. If you can build this loop for a college project, you are already ahead of 90% of developers.