Read this first: One of the most respected engineers in AI research did something quietly shocking in late 2024 — he stopped writing code. Not because he burned out. Because he decided to evolve.
TL;DR: Since December 2024, Andrej Karpathy no longer writes code by hand. Instead, he orchestrates multiple AI Agents running in parallel. His AutoResearch system ran 700 experiments in two days and surfaced 20 real, actionable model improvements — the kind of output that would take a top human engineer two months to produce. This isn't science fiction. It's a work mode shift happening right now.
He's Not Being Lazy. He's Leveling Up.
In December 2024, Andrej Karpathy — former Tesla AI Director, OpenAI founding member, the person who arguably taught an entire generation how to think about deep learning — said something calm and precise in an interview: "I don't write code by hand anymore."
Not because he can't. Because doing so no longer makes sense.
What he does now looks more like conducting: set the direction, define the spec, review the output, and let multiple AI Agents execute in parallel. His vocabulary shifted too. He stopped thinking "which lines of code do I need to write?" and started thinking in macro-actions: What are the boundaries of this task? What does success look like? Where should the Agent be heading?
This isn't just a personal workflow change. It's a signal worth paying attention to.
Figure 1: One person directing multiple AI Agents in parallel — 700 experiments, two days, 20 real improvements
700 Experiments. Two Days.
The clearest way to understand what Karpathy is describing is a single number: 700.
That's how many experiments his AutoResearch system ran in two days. The result? Twenty genuinely useful directions for improving a model.
For context: that output is roughly equivalent to what a strong engineer would produce after two months of intensive manual tuning.
The logic behind AutoResearch isn't complicated. Let the Agent propose a hypothesis, design the experiment, evaluate the result, and iterate — without a human acting as the relay between each step. The loop runs on its own.
But here's the part Karpathy specifically flags as critical: AutoResearch can outpace humans only on verifiable tasks.
What makes a task verifiable? There's a clear standard for right and wrong — did the model accuracy go up or down, did the test pass, what does the metric say? On tasks like these, an AI can run a complete "hypothesize → experiment → validate → improve" loop, and it gets faster and more precise the longer it runs.
Your advantage as a human engineer shifts somewhere else entirely.
Where's Your Moat?
Karpathy's answer is clear-eyed: designing the automation loop itself, and taste, are the real human advantages.
An AI can tell you "this parameter combination improved accuracy by 0.3%." It cannot tell you that you shouldn't be optimizing that metric in the first place. It can find 20 improvements across 700 runs. It cannot tell you which 3 of those are actually worth pursuing. Deciding what questions are worth asking, and what results are worth caring about — that's still your job.
He also offers a counterintuitive debugging rule: when your Agent task fails, assume the problem is yours first, not the AI's.
Specifically, he points the first round of blame toward technical causes: unclear instructions, insufficient context, ambiguous task boundaries. Engineers, he's noticed, tend to immediately question model capability. Most of the time, the real root cause is "your prompt was terrible."
He goes further: treat your prompt instructions like iterable code. Just like you wouldn't ship the first version of a codebase, your Agent instructions need version control, test cases, and continuous revision based on failure patterns.
The Jagged Edge of Model Capability
Karpathy isn't naively optimistic about any of this. He's explicit about a structural limitation in current AI capability, and the metaphor he uses is apt: jagged.
On verifiable, quantifiable tasks, AI performance tends to exceed what you'd expect — it can operate at a speed and scale that humans simply can't match. But on "soft" tasks, strange blind spots appear: aesthetic judgment, strategic tradeoffs, cross-domain intuition, nuanced human communication.
These aren't bugs. They're the current edge of the technology. And understanding where that edge sits is more useful than either blanket trust or blanket skepticism.
He also said something that made a lot of engineers uncomfortable: you need to start writing documentation for Agents, not for humans.
The difference matters more than it sounds. Human-readable docs can be fuzzy, can lean on implied knowledge, can leave gaps for interpretation. Agents don't work that way. They need explicit boundaries, explicit success criteria, explicit error-handling logic. Building that discipline — more than learning any new tool — is what separates people who will actually thrive in this shift from people who will keep wondering why their Agents keep failing.
Three Things You Can Do Right Now
If you're a developer, researcher, or tech lead, Karpathy's observations translate directly into action:
Redefine your unit of work: Stop thinking "which lines do I write?" and start thinking "what are the inputs, outputs, and verification criteria for this task?" Position yourself as the architect, not the craftsperson.
Identify the verifiable parts of your work: Which tasks in your workflow have a clear right/wrong standard? Start there. That's your easiest entry point for building an AutoResearch-style loop.
Treat your prompts like production code: Set up version control for them, log failure cases, iterate continuously. A well-engineered instruction architecture will get you further than switching to a more powerful model.
Closing: The Baton Is in Your Hands
Karpathy's signal isn't "AI is replacing engineers." It's "the engineer's job is moving up a level."
From writing every line, to defining task architecture. From tuning by hand, to designing automated loops. From explaining things to humans, to writing precise instructions for Agents. This isn't a loss of skill — it's a multiplication of leverage.
The real risk is for the people still fixated on every individual line of code — not because they lack ability, but because their eyes are still on the sheet music. They haven't looked up yet to see the orchestra waiting for a conductor.
That orchestra is ready. The baton is in your hands.
References
- Karpathy, Andrej (2025). Insights on building with AI Agents. Public Interview / Conference Talk.
- Karpathy, Andrej (2025). On AI verifiability and the AutoResearch system. Public Interview.
- Karpathy, Andrej (2025). Full interview — not writing code since December 2024. Podcast.
- Karpathy, Andrej (2025). AI anxiety and the decentralized Agent swarm concept. Public Commentary / Social Media.
Found this useful?
Follow for new AI × biomedical research notes:
Or buy me a coffee to keep new content coming.
☕ Buy Me a Coffee