Andrej Karpathy: true AI agents are a decade away

In a new interview, OpenAI cofounder Andrej Karpathy argues truly reliable “agents” are roughly a decade away—thanks to brittle tool-use, long-horizon planning errors, and data issues. Translation: keep building assistants, not autonomy theatre.

The critique in three parts

Cognitive limits: Today’s LLMs struggle with state, credit assignment, and long-horizon decomposition. Error compounds across steps, so multi-tool workflows silently drift off course.
Reinforcement learning reality: RL at agent scale has sparse rewards, high variance, and fragile policies. It works in narrow domains; general business workflows impose too many edge cases.
Data diet and collapse: Heavy reuse of model-generated content risks distribution shift. Without robust retrieval, instrumentation, and fresh human data, agents get dumber where you need them smarter.

What builders should do instead

Assistant UX, not autopilot: Offer capable tools with tight human-in-the-loop, explicit confirmations, and partial automation where the payback is clear.
Deterministic scaffolding: Model calls inside state machines with retries, tool pre/post-conditions, and idempotent side effects. Prefer short horizons, checkpointed plans, and visible state.
Evidence-driven evals: Track success/failure per task template; keep gold-sets; log tool outcomes. Treat agents as systems, not prompts.
RAG + provenance first: Retrieval grounds outputs; provenance keeps lawyers calm. Weigh retrieval quality above raw parameter counts.

Strategy implications

If you’re pitching fully autonomous customer agents, reset expectations to “assistive co-pilots” that accelerate human workflows. Focus roadmaps on reliability, observability, privacy, and on-prem/VPC options. The winners will ship workflows that measurably cut cycle times—not sci-fi demos.