Karpathy Says Stop Coding. A Fastenal Vending Machine Explains Why He’s Right.

THE NUMBER: $230 billion — the current market cap of Cisco, the company that didn’t build websites but built the routing layer that made websites possible. The agent era needs the same thing. Nobody’s building it yet.

Andrej Karpathy — the man who taught a generation of engineers to build neural networks from scratch through Stanford lectures and YouTube tutorials — just told everyone to stop writing code. Manage the agents that write it for you, he said. The guy who wrote the playbook just rewrote it.

This isn’t a theoretical shift. The data is proving it right now. Tomasz Tunguz reports tech hiring is up 16% — but down 45% from peak. Companies are shipping more with fewer people. They’re not hiring new engineers. They’re getting more out of the ones they have by pointing them at agent swarms instead of IDEs. The marginal hire is dead. The marginal manager is everything. Right now, coders scale, managers don’t yet.

Meanwhile, Aaron Levie and François Chollet are mapping what happens when those agents get wallets. Micropayments that never worked for humans — too expensive to subscribe, too low-volume to justify — suddenly make economic sense when an agent can spend $0.10 to access paywalled research mid-workflow. The business models that failed on the human internet are about to work on the agent internet. The commodity pricing of LLMs isn’t a tragedy. It’s the enabling condition — cheap electricity that powers an entirely new economy.

And Perplexity just shipped the most tangible version of this future: an always-on Mac Mini that knows your files, runs while you sleep, and is controllable from any device. It’s not another chatbot. It’s the first dumb terminal on a network that doesn’t exist yet — the same way Netscape was a window into a world that nobody had really built yet.

The trillion-dollar question isn’t who builds the best model. It’s who builds the best router.

Who is the next Cisco?

The Guy Who Wrote the Playbook Just Rewrote It

Andrej Karpathy posted something on Tuesday that racked up 187,000 views in hours. His thesis: the age of the IDE isn’t over — it just looks completely different. The basic unit of programming isn’t a file anymore. It’s an agent. Humans move up a level. You don’t write code. You orchestrate the things that write code.

Yuchen Jin compared it to StarCraft in the replies. That’s not a throwaway analogy — it’s the right one. Real-time strategy. Resource allocation at machine speed. Conflict resolution every few seconds, not every few days at a standup meeting.

Here’s the part nobody’s connecting: there’s a tweet from @mindofachaser going viral this week — 252,000 views — about a construction company that fills its offices with Fastenal vending machines. Drill bits, blades, hardhats — all priced way above Home Depot. The owner doesn’t care. Every minute a worker spends driving to Lowe’s, wandering the aisles, scrolling Instagram in the parking lot — that’s a minute not spent on the job site. The premium on the parts is nothing compared to the cost of the downtime.

That’s the exact shift happening in software right now. The bottleneck isn’t writing code. It’s the friction between needing a thing and having a thing. Between an agent hitting a blocker and a human clearing it. We’re still operating on daily-standup time in a world where agents work at millisecond speed. The cost isn’t the tool. It’s the wait.

Naval Ravikant put it in eight words: “AI is going to drain a lot of moats.” A Florida homeowner just proved it at street level — sold his house using ChatGPT instead of a real estate agent. No staging. No marketing team. No 6% commission. Real estate agents’ moat was always information asymmetry dressed up as expertise. That moat just drained.

But here’s a Charlie Munger nuance that the “moats are dead” absolutists will likely miss: AI drains moats built on information asymmetry and labor arbitrage. It does not drain moats built on trust, regulatory capture, or network effects. Your realtor is vulnerable. Your surgeon is not (at least until Elon gets his Optimus surgical team working). The question every executive needs to ask: which kind of moat is mine?

What business leaders need to know: The skill in shortest supply isn’t coding — it’s directing. If your org still measures engineering productivity by lines of code shipped, you’re measuring the wrong century. The conversation to have with your team this week: “If our engineers managed twelve agents instead of writing code themselves, what would we need to change about how we run projects?” The companies that figure out orchestration-at-speed will ship what their competitors can’t even staff for.

Who Builds the Cisco for Agents?

Aaron Levie’s thread on agent economics laid out a future that most people aren’t thinking about yet. When trillions of agents use the internet, entire categories of business models that failed for humans start working. Paywalled research data that no individual would pay $1,000 to access? An agent with a budget pays $0.10 mid-workflow and moves on. APIs and tools that couldn’t sustain subscriptions? Agents interact on a per-transaction basis at fractions of a cent.

Here’s where it gets real. Think about LexisNexis, the New York Times, the Financial Times, the Economist — companies sitting on decades of paywalled intelligence. Right now, an individual subscribes for hundreds of dollars a year or hits a wall. Most people bounce. But an agent running a research workflow doesn’t need a subscription. It needs one article, right now, for a penny. A cent per article consumed agentically. Maybe it costs you a dollar a day in micro-fees — and your product gets meaningfully better. The publisher already has the information. Exposing it at micro-scale doesn’t cannibalize the $500 annual subscription because the agent isn’t trying to recreate the newspaper for a dollar. It’s looking for one specific signal in one specific article. Everyone wins. The infrastructure to make this work is what’s missing.

François Chollet called the timeline: 1-2 years before we see this at scale. That’s not a prediction about AI capability. It’s a prediction about economic infrastructure.

The analogy everyone should be thinking about isn’t software. It’s the internet itself. Before the web could scale, someone had to build the routing layer — the collision detection, the packet switching, the air traffic control that made sure data got where it needed to go without crashing into other data. Cisco didn’t build the websites. Cisco built the thing that made websites possible. That company is worth $230 billion.

The agent era needs its Cisco. Not another model. Not another chatbot. A routing layer for tasks. Something that handles real-time conflict resolution when Agent A needs a resource that Agent B is using. Priority queuing when twelve agents hit the same API simultaneously. Dynamic reallocation when a workflow bottleneck appears — not at the next standup, but in milliseconds. The Fastenal vending machine eliminated friction between “I need a drill bit” and “I have a drill bit.” The agent router eliminates friction between “I need this data” and “I have this data” at software speed.

Perplexity’s always-on Mac Mini is the first hardware that takes this seriously. It’s not a chatbot that runs locally. It’s a machine that never turns off, knows your files, manages your sessions, and works while you sleep — controllable from any device on earth. For a small business owner running constantly-changing digital marketing or a data-intensive operation, this is the form factor that makes agentic orchestration tangible. Not a tab you open when you need help. Infrastructure that’s always running.

And here’s where it gets interesting: there’s no reason your iPhone can’t evolve into this. Enough memory, enough compute, always connected — your phone becomes an always-on orchestration node. The smartphone didn’t just make calls mobile. It made the internet mobile. The always-on agent device doesn’t just make AI portable. It makes the agent economy portable.

Microsoft clearly sees this — which is why their new $99/month E7 tier bundles Copilot and Agent 365, and runs on Anthropic’s Claude, not just OpenAI. The irony is almost too perfect: the company that tried to own the AI layer is now licensing someone else’s intelligence and charging rent on the bundle. They don’t care which model wins. They just want to be the landlord. We’ve seen this movie before — Microsoft didn’t build the internet’s routing layer either. They ran on top of it.

Here’s a thought nobody’s having: maybe the prototype for all of this already exists — in human form. Say what you want about Elon Musk, but strip away the tweets, the politics, the main-character energy, and look at how the man actually operates. No food at meetings. No small talk. Walk in, identify the bottleneck, clear it, move on. He runs five companies simultaneously not because he’s smarter than everyone else, but because he operates as a human orchestration router by default. He doesn’t do the engineering. He clears the path for the engineers and makes the calls they can’t make fast enough. Every meeting is a conflict-resolution loop. Every decision is a resource-reallocation event. He’s been running the Cisco protocol — routing tasks, resolving collisions, managing bandwidth across parallel operations — his entire career. Not many people on earth operate this way. Perhaps that’s the rarest skill of the next decade: the ability to think like a router, not an endpoint. The agent economy doesn’t need more coders. It needs more people who can do what Musk does — but with machines instead of humans, at machine speed.

The action item: Stop asking “which AI model should we use?” Start asking “who’s building the infrastructure that lets agents find, pay for, and use what they need — automatically?” The model is the commodity. The router is the moat. If you’re allocating capital, look at the companies building agent infrastructure — discovery, payments, orchestration — not the ones burning billions on training runs. Bet on the Fastenal, not the drill bit.

What This Means For You

Two forces are converging, and they move in the same direction. The humans are moving up — from coders to conductors, from individual contributors to orchestrators of machine fleets. And the machines are moving out — from tools you open at your desk to infrastructure that works while you sleep, spends money on your behalf, and routes around obstacles in real time.

The orchestration skill is the new literacy. Three years ago the premium was on prompt engineering. Today it’s on managing agents at scale — clearing bottlenecks in real time, maintaining context across a dozen parallel workflows, knowing when to let the machine run and when to intervene. If you’re not building this muscle in your organization, your competitors are.

Bet on infrastructure, not intelligence. The trillions being spent on model training will commoditize exactly the way compute always does. The money gets made on what runs on top of the commodity layer — routing, discovery, orchestration, payments. The Cisco of agents doesn’t exist yet. When you find it, that’s your position.

Audit your moats this quarter. If your competitive advantage depends on information asymmetry, labor arbitrage, or pricing opacity, AI is coming for you faster than you think. If it depends on trust, regulatory position, or network effects, you have time — but not as much as you’d like. Be honest about which category you’re in.

The winners of 2027 won’t be the companies with the best models. They’ll be the ones who figured out how to conduct the orchestra.

Three Questions We Think You Should Be Asking Yourself

If my best engineer became a general contractor tomorrow — managing twelve agents instead of writing code — what breaks in how we plan, review, and ship? Most orgs are structured around human-speed execution cycles. Daily standups, two-week sprints, quarterly planning. None of that maps to a world where agents ship in hours and bottlenecks need clearing in seconds. The org chart hasn’t caught up.

What data does my business sit on that’s worth a penny per query to an agent — and how would I even expose it? You might be LexisNexis and not know it. Every company has proprietary data that’s too expensive to package for human subscribers but perfectly priced for agentic micro-consumption. The infrastructure doesn’t exist yet to monetize this. But the companies thinking about it now will be ready when it does.

Is my competitive moat built on something AI can replicate — or something it can’t? Information asymmetry, labor cost advantages, and pricing opacity are all on a timer. Trust, regulatory position, brand, and network effects still hold. Be brutally honest about which category yours falls in. Then act accordingly — because the drain is faster than anyone’s forecasting.

---

The age of the IDE isn’t over. We’re going to need a bigger IDE. It just looks very different because humans now move upwards and program at a higher level — the basic unit of interest is not one file but one agent. It’s still programming.”

— Andrej Karpathy

---

— Harry and Anthony

Sources

• Andrej Karpathy on the future of programming
• Aaron Levie on agent economics and micropayments
• François Chollet on agents as economic actors
• Tuki on Karpathy’s playbook rewrite
• Tuki on Naval and moat drainage
• @mindofachaser on Fastenal vending machines at construction companies
• Florida homeowner sells house using ChatGPT
• Perplexity’s always-on Mac Mini
• Tomasz Tunguz on tech hiring trends