A year ago, AI coding assistants autocompleted your lines. Now they write entire features while you watch. Andrej Karpathy thinks that shift is just the start.
In a new talk titled "The End of Coding," Karpathy lays out his case that traditional programming is being replaced by something fundamentally different: AI agents that run in loops, writing code, executing it, checking the output, and iterating until the result works. He calls this the "loopy era" of AI.
What "Loopy" Actually Means in Practice
The term sounds whimsical, but it describes a real pattern anyone using Claude Code, Cursor, or Aider has already experienced. Instead of a human writing code line by line, you describe what you want. The agent writes a draft, runs it, reads the error message, fixes the code, runs it again, and repeats. The human becomes a reviewer and director rather than a typist.
Karpathy's framing goes further than coding assistants, though. He introduces the concept of "AutoResearch" - AI agents that don't just write code but run entire research workflows autonomously. Think: an agent that formulates a hypothesis, writes the experiment code, runs it, analyzes results, and proposes next steps. The human sets the goal and evaluates the output.
This is not a distant prediction. Tools like Claude Code already operate in agent loops by default - they propose changes, run tests, read failures, and fix issues in a cycle. Devin, Codex, and similar tools do the same. The infrastructure for loopy coding exists today.
The Part That Matters for Non-Engineers
Karpathy is not just talking to developers. The implication of agent loops is that building software stops being a specialist skill. If an AI agent can take a plain-English description, write working code, test it, and ship it - the bottleneck moves from "can you code" to "can you describe what you want clearly."
We have already seen this with tools like Bolt, Lovable, and Replit's agent mode, where non-technical users build working apps by describing them. Karpathy's argument is that this is the permanent direction, not a gimmick.
Where the Skepticism Should Land
The "end of coding" framing is provocative by design, and Karpathy knows it. Coding agents still fail on complex, multi-file refactors. They hallucinate APIs that do not exist. They struggle with codebases larger than their context window (the amount of text they can process at once). Senior engineers still spend significant time fixing what agents produce.
But the trajectory is clear. A year ago, these agents could barely write a function. Now they build features across multiple files, run their own tests, and self-correct. The gap between "agent-assisted coding" and "agent-does-the-coding" is shrinking fast.
Karpathy's track record gives this weight. He co-founded OpenAI, led Tesla's Autopilot AI team, and built some of the most-watched AI education content online through his Eureka Labs project. When he says coding as we know it is ending, that is a data point, not just an opinion.