Beyond the Cursor: A Primer on Agentic Coding

Columbia University — Feb 11, 2026

These are the core ideas from a presentation I gave at Columbia on how AI coding tools have evolved and where they’re heading.

The Three Generations

• Gen 1: Autocomplete. IntelliSense, Tabnine. Mirrors your keystrokes. Reduces boilerplate but has no concept of the system.
• Gen 2: Copilots. ChatGPT, Gemini. Understands intent behind syntax. Suggests structural patterns, not just completions.
• Gen 3: Agents. Claude Code, Gemini CLI, Codex. Executes based on intent. The model drives shell primitives directly — the cursor disappears.

We’ve moved from “drawing the code” to “defining the logic.”

What Agency Actually Means

• The Loop: Plan → Act → Observe → Iterate. Unlike Gen 2 where the loop ends at the output window, Gen 3 agents observe terminal output and compiler errors to course-correct autonomously.
• Equilibrium: The agent’s goal is returning the system to a stable state (green tests, managed debt) after the disruption of a new feature request.
• Stress: Intentional friction — human-in-the-loop checkpoints — to prevent the agent from drifting into hallucinated solutions.

The Takeaway

The AI provides the structure. You provide the intent. The job isn’t to automate everything — it’s to curate the environment so the automation stays on track.