Andrej Karpathy hasn't written a line of code since December. The shift to 'agentic engineering' will change how your next business tool gets built.
Article text
One of the best programmers alive hasn't typed a line of code since December.
Andrej Karpathy built Tesla's self-driving AI. He taught Stanford's most popular deep learning course. He co-founded OpenAI. If anyone has the chops to write code by hand, it's him.
And he stopped. Not because he retired. Because the tools got good enough that directing AI agents to write code is now faster, better, and more productive than doing it himself.
His words: "I don't think I've typed like a line of code probably since December, basically, which is an extremely large change." He described the shift as a "magnitude 9 earthquake" to how programming works.
If you're a small business owner, you don't care about Karpathy's daily workflow. But you should care about what happens when the people building your software suddenly get 3-5x more productive. Because that's what's happening right now.
Wondering what this means for a project you've been putting off? We do free 30-minute discovery calls where we help you figure out whether now is the right time to build.
From Writing Code to Directing Agents
Here's what Karpathy's workflow looks like now, as he described it : he sits in front of a tiled monitor with multiple AI coding agents running in parallel. He assigns each agent a feature or task using plain English. They write the code. He reviews the output, catches errors, and routes the next set of tasks.
He calls this "agentic engineering." The developer becomes a project manager directing a team of AI agents, rather than a craftsperson typing code line by line.
The numbers tell the story. In September 2025, Karpathy wrote 80% of his code himself, with AI handling 20%. By December, that ratio flipped. Now he says he's "mostly programming in English, a bit sheepishly telling the LLM" what to build.
This isn't a fringe experiment. This is the trajectory the entire industry is on.
Why This Matters If You Run a 30-Person Company
Let me translate this into business terms.
Custom software is about to get meaningfully cheaper.
When a developer can direct AI agents to handle the repetitive parts of building software, they can finish projects faster. A dashboard that took 120 hours to build six months ago might take 40-60 hours today. The developer still needs expertise to architect the solution, review the code, and handle edge cases. But the raw production time is compressing fast.
We've seen this firsthand. A commercial cleaning company came to us last month wanting an automated scheduling and route optimization tool. Twelve months ago, a project like that would have been a 6-8 week build. We scoped it at 3 weeks. Not because we cut corners, but because the building process itself has changed. Our team used AI coding agents for roughly 60% of the implementation, focusing their own time on the business logic and integration points that AI still gets wrong.
The skill that matters is knowing what to build, not how to type it.
This is a subtle but important shift. When AI handles the typing, the bottleneck moves upstream. The hard part becomes understanding your business process well enough to describe it clearly, then catching the mistakes AI makes before they ship.
Karpathy himself flagged this: AI agents make "subtle conceptual errors that a slightly sloppy, hasty junior dev might do." They make unfounded assumptions. They over-complicate solutions. They generate 1,000 lines of code when 100 would do.
The developer's value has shifted from production to judgment. And for business owners hiring developers or consultants, this changes what you should look for.
What to Ask Your Developer (or AI Consultant)
If you're hiring someone to build custom tools or automation for your business, the Karpathy shift changes the conversation. Here's what to ask:
"Are you using AI coding agents? How?"
This is no longer a nice-to-have. A developer or consultancy that isn't using AI agents in 2026 is billing you for hours they don't need to spend. You want to hear specifics: which tools they use, how they review AI-generated code, and how it affects their project timelines and pricing.
"How do you catch the mistakes AI makes?"
This is the critical question. AI agents write functional code quickly, but they introduce subtle bugs. A good team has a review process: automated testing, human code review of AI output, and specific checks for the kinds of errors AI tends to make (over-engineering, dead code, wrong assumptions). If they just let the AI run and ship whatever comes out, that's a red flag.
"Has this changed your pricing?"
If a team's productivity has doubled or tripled thanks to AI tools, that should show up in their quotes. Not all of the savings will pass through to you (teams invest the time savings into higher-quality work and more thorough testing). But if someone's charging 2024 rates for 2026 productivity, ask why.
"What parts does AI handle vs. what do humans handle?"
You want to hear that AI handles the repetitive code generation while humans handle architecture, business logic, integration design, and quality assurance. If the answer is "AI does everything," walk away. If the answer is "we don't use AI," also consider walking away. The sweet spot is a team that knows exactly where the boundary is.
The "Slopacolypse" Risk
Karpathy coined a term worth knowing: the "slopacolypse." He predicts 2026 will see a flood of "almost right, but not quite" AI-generated code across the internet. Code that works in demos but breaks under real conditions. Code that's technically functional but poorly structured and expensive to maintain.
This matters for business owners because it creates a quality gap. The consultants who know how to use AI agents well will deliver faster AND better. The ones who use AI tools carelessly will deliver faster but worse. And it's hard to tell the difference until something breaks.
Here's how to protect yourself: ask for a maintenance plan. Ask what happens when something needs to change six months from now. AI-generated code is easy to create but can be expensive to modify if it wasn't structured thoughtfully. The team that builds your tool should be the team that can explain every piece of it, even the parts AI wrote.
What This Looks Like in Practice
Think of it like hiring a general contractor for a home renovation.
Ten years ago, the contractor and their crew did everything by hand. Measuring, cutting, fitting, finishing. The skill was in the hands.
Today, a good contractor uses CNC machines, laser levels, and prefab components. They still need to understand structural engineering, building codes, and material properties. They still need to catch when something doesn't look right. But the physical production is faster because the tools are better.
The contractors who get the best results aren't the ones who refuse to use power tools. And they're not the ones who let the machines run unsupervised. They're the ones who know exactly when to use the tool and when to use their own judgment.
That's where software development is right now. The tool (AI agents) handles the production. The developer handles the thinking. And the businesses that benefit most are the ones who find teams that have figured out that balance.
The Window Is Open
Here's the practical takeaway: if you've been putting off building a custom tool, automation, or internal system because the quote was too high or the timeline too long, check again. The economics of software development shifted meaningfully in the last three months. Projects that were $50,000 six months ago might be $25,000-35,000 today. Timelines that were 8 weeks might be 3-4.
This won't last forever. As the market adjusts, pricing will stabilize at a new normal. But right now, there's a gap between what things cost to build and what some firms are still charging. The teams that have adopted agentic engineering are delivering more for less. The ones that haven't are still quoting last year's rates.
The question isn't whether this shift is real. Karpathy, with his credentials, called it the biggest change to programming he's seen. The question is whether you take advantage of the window while the economics are in your favor.
Got a project you've been sitting on? Book a free workflow call and we'll give you an honest assessment of what it would cost today vs. what you might have been quoted before. No commitment. Just a current market read.