The Decade of AI Agents: Andrej Karpathy’s Vision for AGI

Andrej Karpathy, one of the most respected minds in artificial intelligence, recently set the tone for the next decade of AI development with a phrase that’s now resonating across the tech world: “This isn’t the year of agents — it’s the decade of agents.” His statement, made during a long-form discussion on the Dwarkesh Podcast, offers a crucial reality check for an industry riding high on the promise of autonomous AI systems.

Why It’s Not the “Year of Agents”

The AI community dubbed 2025 the “year of agents,” pointing to rapid progress in LLM-powered tools that claim to perform complex, multi-step tasks with minimal user interaction. Yet, Karpathy argues that most of these systems are prototypes at best. They can execute impressive demos but fall dramatically short when tested against real-world reliability and cognitive competence.

In his words: “They just don’t work. They don’t have enough intelligence, they’re not multimodal enough, they can’t do computer use, and they don’t have continual learning.” He adds that achieving true agentic capability will take close to a decade of dedicated research and iteration.​

What Makes True AI Agents So Hard

1. Cognitive Deficits

Current AI agents, even the most advanced ones like Claude or Codex, lack fundamental cognitive depth. They can generate coherent responses but often fail at tasks requiring reasoning, adaptation, and precision. Karpathy likens a true AI agent to “an intern you could actually hire” — one that understands context, remembers lessons, and can independently plan work.​

2. Lack of Memory and Adaptation

Perhaps the largest gap lies in memory. Modern systems cannot retain contextual knowledge across tasks. Every new session is a reboot — there’s no sense of experience, continuity, or the ability to improve over time. Karpathy contrasts this sharply with how humans learn continuously, integrating new experiences into long-term understanding.​

3. Reinforcement Learning Bottlenecks

Karpathy also criticizes reinforcement learning as “sucking supervision through a straw.” Agents only receive a single reward signal after completing a task, making it nearly impossible to ascribe credit or blame to specific actions. This coarse feedback loop creates brittle agents that often reinforce poor strategies as readily as good ones.​

The Vision: A Decade of Gradual Maturity

Karpathy’s prediction that it will take a decade to “work through all of these issues” isn’t pessimistic — it’s realistic. True agentic systems must learn to think, remember, and act over long time horizons. This evolution will happen through:​

• Building multimodal systems that reason over text, vision, and code simultaneously

• Developing continual learning architectures that preserve context across sessions

• Creating reliable control loops to supervise complex tasks safely and effectively

As Karpathy puts it, the real goal is to reach a point where AI agents act as dependable employees rather than clever tools — partners that can be trusted to execute work autonomously under human guidance.​

The Broader Consequences: Work, Governance, and Patience

Karpathy’s framing of the “decade of agents” extends beyond technical nuance. It represents a broader call for patience, humility, and engineering maturity. While investors and media chase fast breakthroughs, Karpathy advocates for sustained progress grounded in continuous refinement — the same patience that preceded the deep learning explosion a decade earlier.​

His outlook also comes with a sober social implication: humans won’t become obsolete, but their roles will shift from execution to orchestration. Future organizations will manage “fleets of AI agents” — systems that run operations, optimize processes, and report to human supervisors overseeing multiple intelligent programs simultaneously.​

Ten Years to Build Trustworthy Intelligence

Karpathy concludes that the true test of this decade isn’t building flashier agents but building trustworthy ones. Smart enough to help, stable enough to rely on, and safe enough to deploy at scale.

He sums it up sharply: “We’ll be working with these things for a decade. They’re going to get better, and it’s going to be wonderful.” The future of AI agents will indeed be transformative — but only if the world gives it the decade it deserves.