Introduction
AI agents are one of the hottest trends in 2024, but a recent statement from Andrej Karpathy, OpenAI co-founder, has thrown cold water on the hype: AI agents don't work yet and will take at least a decade before they're truly autonomous. This isn't skepticism from an outsider, but the assessment of someone actually building frontier technology.
What Karpathy Said About AI Agents
AI agents lack sufficient intelligence to operate autonomously, according to Karpathy's recent Dwarkesh Podcast appearance. His words are direct: "They're cognitively lacking and it's just not working. It will take about a decade to work through all of those issues."
The identified problems are multiple and structural: agents lack adequate multimodal capabilities, can't use computers like humans do, miss continuous learning, and crucially, cannot retain information between sessions.
"They don't have enough intelligence, they're not multimodal enough, they can't do computer use and all this stuff. They don't have continual learning. You can't just tell them something and they'll remember it."
Andrej Karpathy, OpenAI Co-founder
The Gap Between Hype and Market Reality
While companies invest billions in AI agent platforms, concrete data tells a very different story:
- 95% of AI pilot projects failed to meet stated objectives according to Gartner research
- 50% of organizations that expected to significantly reduce customer service headcount by 2027 are abandoning those plans
- Startups that invested in $500K agent platforms aren't seeing meaningful returns yet
This creates a real problem for enterprises: many have already purchased promising solutions, but the promised autonomy isn't yet available.
When AI Agents Will Actually Work
According to Karpathy, truly autonomous AI agents will arrive around 2035. This isn't arbitrary: the timeline aligns perfectly with global Wi-Fi 8 deployment and the evolution of global hardware infrastructure.
The issue isn't just software. Current infrastructure isn't designed to handle the low-latency, symmetric traffic that autonomous agents would generate. That's why we're still in infrastructure buildout phase: data centers, edge computing, and connectivity standards must evolve together.
The Model That Works Today: Assisted Automation
While we await true agents, some companies have found an approach that works. McKinsey built an AI agent using Microsoft Copilot Studio that monitors email for incoming project proposals. Result: review time dropped from 20 days to 2.
But there's a critical caveat: a human must verify what the agent produces. It's not full automation, it's assisted automation. And it works precisely because it doesn't promise total autonomy.
Why Assisted Automation Wins
This model also reflects Karpathy's vision: agents should function like "an employee or intern you'd hire to work with you." But with one crucial difference: interns learn; current agents don't. They can't retain information between sessions, severely limiting their usefulness for complex tasks.
What to Expect in the Coming Years
The reality is less glamorous than the hype, but more sustainable:
- 2024-2026: continued assisted automation; companies optimize processes with supervised agents
- 2026-2030: incremental evolution of intelligence and multimodality; specific use cases improve
- 2030-2035: completely renewed infrastructure; first truly autonomous agents emerge in limited domains
- Post 2035: true agents capable of continuous learning and memory retention
Startups and companies promising complete automation by 2027 are selling a product that doesn't exist yet and won't for quite some time.
The Necessary Mindset Shift
As Karpathy wrote on X, the true paradigm shift is this: "I want AI to make fewer assumptions and ask/collaborate with me when not sure about something. I want to learn along the way and become better as a programmer, not just get served mountains of code that I'm told works."
This mindset separates companies that will succeed from those that will burn cash:
- Winners: build for augmentation, not replacement
- Losers: promise complete automation in the next 2-3 years
AI agents are tools that make humans better, not substitutes. This is the reality companies should build their technology stack on.
FAQ
When will AI agents actually work?
According to Andrej Karpathy, OpenAI co-founder, AI agents will achieve true autonomy around 2035. This timeline aligns with global infrastructure evolution, not just software development.
Why don't current AI agents work?
AI agents lack sufficient intelligence, aren't adequately multimodal, can't use computers like humans do, and crucially, cannot retain information between sessions. They're cognitively limited and architecturally incomplete.
What percentage of AI projects have failed?
According to Gartner, 95% of companies that implemented AI pilots encountered failures in meeting stated objectives.
Is there an AI agent model that works today?
Yes: assisted automation. Companies like McKinsey use AI agents to accelerate processes, but humans always verify results. It's collaboration, not total autonomy.
What are the main limitations of today's AI agents?
Lack of continuous learning, inability to retain memory between sessions, insufficient intelligence, limited multimodal capabilities, and inability to use computers like humans do.
What makes total AI agent autonomy impossible by 2027?
It's not just software. Global infrastructure (data centers, Wi-Fi, edge computing) must evolve alongside software. We're in infrastructure buildout phase, not autonomous deployment yet.
How should companies approach AI agents today?
They should pursue assisted automation solutions that enhance human productivity, not those promising total replacement. Agents should collaborate with humans, not replace them.