The buzz for AI agents is at fever pitch as investors are proclaiming 2025 as “the year of the agent.” But the founders of OpenAI and one of the leading voices in artificial intelligence, Andrej Karpathy, is raising a warning flag.
In a candid review of current AI agent technology last week on the Dwarkesh Podcast, Karpathy passed judgment: They already don’t perform well enough. And he believes it’s going to take well over a decade to fix fundamental problems that are inhibiting them.
“They just don’t work. They don’t have enough intelligence, they’re not multimodal enough, they can’t do computer use and all this stuff,” said Karpathy, who now runs Eureka Labs, an AI-native educational startup. “They don’t have continual learning. You can’t just tell them something and they’ll remember it. They’re cognitively lacking and it’s just not working.”
Former OpenAI Co-founder Andrej Karpathy, The Agent Hype and the Call for Cooperative AI
Agents are basically computerized assistants who are responsible for executing tasks on their own.
As opposed to chatbots that can simply respond to inputs, agents are responsible for breaking down complex problems, drawing up a plan of action, and executing tasks without constant human checking and interference. Consider them as computerized employees who are capable of coding projects as well as office jobs.
He is concerned with more than technical constraints. He fears that the AI community is creating technologies for a future that does not yet exist, one in which complete autonomous AI technologies take control of everything while humans are on the bench.
The Call for Cooperative, Explainable AI and the Error Compounding Problem
Known for his rapid-fire speaking style, Karpathy later clarified his thoughts in a follow-up post on X. “My critique of the industry is more in overshooting the tooling w.r.t. present capability,” he wrote. “The industry lives in a future where fully autonomous entities collaborate in parallel to write all the code and humans are useless.”
He envisions a cooperative method in which humans and machines share on a task. He is interested in AI tools that explain their workings, defend their thought processes, and solicit when in doubt. The goal is not to replace human expertise but to complement it.
“I’d like it to pull in API docs and demonstrate that it used it correctly. I’d like it to make fewer assumptions and query/ask/share with me when it’s not quite sure of something. I’d like to see it learn over the course of it and make you a better programmer yourself rather than be presented with stacks of code that you’re simply informed works,” he explained.
He is not alone in his skepticism. In a blog post on LinkedIn last year, ScaleAI growth lead Quintin Au highlighted another significant weakness of current AI agents: the error compounding problem.
The Case for Pragmatism and Patience in Developing Reliable AI Agents
In Au’s estimates, big language models perform with approximately 20% error on single actions. Sounds tolerable? The arithmetic gets nasty quickly. If you’re going to perform five consecutive tasks as an agent, your chance of receiving each step correctly falls to 32%. The bigger the task, the greater its chance of going awry.
This critical limitation makes current AI agents inappropriate for a majority of practical applications since situations require accuracy and consistency.
Amongst Karpathy’s bigger fears is that of spreading so-called AI “slop”, weak content created with AI programs. If we create agents that are capable of churning out gargantuan amounts of code or content without substantial human review, we are taking a chance on drowning the digital world with subpar, error-prone output.
The problem isn’t just about quality. It’s about losing the learning process. When AI does everything for us, we stop developing our own skills and understanding. Programmers don’t improve if they’re just rubber-stamping AI-generated code they don’t fully comprehend.
Despite his critique, Karpathy made clear he’s not an AI skeptic. He believes the technology will eventually get there; it just needs more time.
“My timescales for AI are 5-10 times more glum than what you’d expect in your neighborhood SF AI house party or on your Twitter timeline, but relatively optimistic relative to a growing tide of AI skeptics and deniers,” he continued.
His point is actually patience and pragmatism. The AI industry must aim to develop tools that are reliable today instead of overpromising abilities that won’t be developed for several years. It is preferable to develop truly valuable collaboration tools instead of harboring hopes of complete automation before technology is mature enough.
Humans are still at the center for now. And that is exactly where they should be, according to Karpathy.
