Why Karpathy's 'decade of agents' warning is back on X

A clip of Andrej Karpathy is making the rounds on X again this summer, and it is worth a minute if anyone has pitched you on AI "agents." Karpathy is the former Tesla AI director and OpenAI cofounder who coined the term "vibe coding," so he is not a bystander. The version getting reshared, captured by @karlmehta, quotes Karpathy saying "this is the decade of agents." His point is blunt: 2025 was not the year of agents, and treating it that way sets everyone up for disappointment.

Where the warning came from

The quote traces back to an interview on the Dwarkesh Patel podcast in October 2025, documented at the time by Simon Willison on his weblog. Karpathy's concern grew out of his self-driving years at Tesla, where a slick demo was always easy and the reliable product was always hard. Willison quotes him saying the remaining problems "will take about a decade" to work through.

He was not calling agents useless. He was saying the tooling and the marketing are running years ahead of what the models can actually do today. That distinction is the whole ballgame.

Why his take carries weight

Here is what makes Karpathy credible rather than cynical: he is a heavy user, not a skeptic. In a separate post about his own coding routine, @karpathy described flipping to "80% agent coding and 20% edits+touchups." He leans on this stuff every day.

So when the same person who codes almost entirely through agents tells you the fully autonomous version is a decade out, that is not a hater talking. That is a power user drawing a careful line between what works now and what is still a demo. It is why the reaction on X has been a steady chorus of @DataChaz and others posting "Karpathy was right" as every lab races to ship agents this year.

What "agent" actually means when someone sells you one

Part of the confusion is that the word "agent" has been stretched to cover two very different products:

• A narrow, supervised assistant that drafts a reply, books a slot, or follows up on a lead and then hands off to a human. This works well today.
• A fully autonomous "employee" that runs an entire workflow end to end with no oversight. This is mostly still a demo.

Karpathy's warning lives in the gap between those two. That gap is exactly where budgets get burned: a business buys the second thing, gets the first thing, and concludes AI does not work.

What this signals if you run a business

The lesson is not "wait a decade." It is to buy the version that works now and ignore the version that is still a demo.

The AI that reliably answers your phone after hours, follows up the moment a lead goes cold, fills a canceled appointment, or drafts a quote for a human to approve: that is real, it is here today, and it pays for itself in weeks. The AI that supposedly replaces your entire front office with zero human oversight: treat that pitch the way Karpathy treats "the year of agents," with concern and a lot of questions.

The businesses actually winning with AI right now are not the ones chasing full autonomy. They are the ones automating one painful, well-defined task at a time, keeping a person in the loop wherever judgment matters, and expanding from there once the first thing is boring and dependable.

That is exactly how we scope work at New Face Design. Our free process audit looks at where you are losing hours or leads, picks the one or two tasks worth automating first, and skips the shiny ones that are still demos. You get an honest read on what is ready and what is not, which, funny enough, is the same thing Karpathy has been trying to tell the whole industry.