Building AI Agents with a True Cognitive Core

This article summarizes and builds on key ideas from "We're summoning ghosts, not building animals" by Andrej Karpathy (Tesla, OpenAI), published October 17, 2025. Original content: YouTube interview with Dwarkesh Patel.

Understanding the true requirements for building capable AI agents has become critical as organizations worldwide invest billions in automation initiatives. Andrej Karpathy (Tesla, OpenAI) offers a sobering perspective on why current approaches fall short and what's needed for genuine progress. His insights challenge the prevailing narrative of imminent AI autonomy and provide a roadmap for more realistic, foundational development that resonates deeply with enterprise AI strategies focused on sustainable, scalable transformation rather than flashy demonstrations.

The path to capable AI agents is proving longer than many expected. While large language models have unlocked impressive abilities, they still lack the integration, adaptability, and autonomy needed to function as reliable digital workers. Leaders aiming to deploy AI in meaningful ways must confront the gap between today’s brittle systems and the resilient, multi-modal agents envisioned in roadmaps.

The next decade will be shaped by gradual layering of capabilities, not sudden leaps to full autonomy. Real progress will come from refining architectures and training methods that produce a strong cognitive core—an engine for reasoning, learning, and generalization—before attempting to match the flexible intelligence seen in biological systems.

Main Story

Attempts to build agents too early have repeatedly run into hard limits. Early reinforcement learning on games or web tasks failed because reward signals were sparse and representation learning immature. Modern agents start with rich pretrained representations from large language models, then add action-taking capabilities. This pretraining acts as a kind of “crappy evolution,” bootstrapping intelligence from internet-scale imitation rather than biology.

We're not building animals. We're building ghosts… by imitating internet documents. — Andrej Karpathy

Pretraining yields embedded knowledge and emergent skills such as in-context learning, but it also encourages memorization over adaptability. Models often overfit to the training data manifold, struggling with novel or precision-critical tasks. A leaner, algorithm-rich cognitive core—less burdened by rote knowledge—may generalize better.

Reinforcement learning remains a blunt tool for intelligence tasks. Outcome-based weighting amplifies noise by upweighting entire trajectories. Process-based supervision, where feedback is given at each step, offers promise but is vulnerable to adversarial exploitation of LLM judges. Synthetic data generation faces its own challenge: model collapse, where outputs lose diversity and reduce learning potential.

Human cognition offers useful contrasts. Forgetting forces generalization, while offline consolidation distills experience into lasting capabilities. Current AI lacks a continual learning mechanism that allows long-term personalization without retraining from scratch.

Code generation illustrates the gap between hype and reality. Autocomplete can accelerate boilerplate and familiar patterns, but autonomous handling of novel, architecture-level changes is still out of reach. Without human scaffolding, agents misfire when tasks deviate from common internet examples.

Technical Considerations

For engineering leaders, several constraints define the near-term landscape:

• Memory limitations: Models cannot sustain working memory across sessions without external systems, making persistent personalization difficult
• Modality integration: Text, image, audio, and structured data capabilities remain siloed; fluid transitions are rare and brittle
• Tool use: Complex computer interactions require human scaffolding; autonomous execution is unreliable in edge cases
• Training trade-offs: Reducing memorization may improve generalization but risks losing useful embedded knowledge
• RL pitfalls: Sparse rewards and noisy trajectory weighting make reinforcement learning inefficient for cognitive skill acquisition
• Synthetic data risks: Model collapse reduces entropy, limiting the diversity needed for robust learning

Leaders should treat these as design constraints when planning agent architectures, integration paths, and long-term capability roadmaps.

Business Impact & Strategy

From a business perspective, the reality of incremental progress changes timelines and investment strategy:

• Time-to-value: Deployments will deliver narrow wins before broad autonomy; expect multi-year horizons for general-purpose agents
• Cost vectors: Ongoing iteration and retraining to maintain adaptability will drive OPEX; memory and modality extensions add infra complexity
• KPIs: Track reliability in narrow domains, diversity of outputs, and successful tool integrations rather than generic “intelligence” scores
• Org design: Pair AI agents with human oversight in hybrid workflows; use them to augment, not replace, skilled staff in high-stakes contexts
• Risk mitigation: Stress-test process supervision and synthetic data pipelines for adversarial exploits before scaling

These realities suggest a staged adoption plan—starting with contained, high-use use cases and evolving toward broader integration.

Key Insights

• Pretraining provides a rich base but encourages memorization over adaptability
• A lean cognitive core may improve reasoning and generalization
• Reinforcement learning is noisy and inefficient for complex cognitive skills
• Process-based supervision needs safeguards against adversarial manipulation
• Synthetic data pipelines risk collapse without entropy-preserving interventions
• Human cognition’s forgetting and offline consolidation could inspire new continual learning mechanisms

Why It Matters

Without a robust cognitive core, AI agents will remain brittle and narrow. For technical leaders, this means focusing on foundational architecture and training methods rather than chasing premature autonomy. For business leaders, it means setting realistic timelines, budgeting for iterative improvement, and designing workflows that blend AI assistance with human judgment. Progress will be measured in sustained capability gains—not in sudden breakthroughs.

Actionable Playbook

• Target the cognitive core: Reduce rote memorization in training to force reliance on reasoning; success = improved performance on novel tasks
• Prototype process supervision: Build domain-specific feedback loops with intermediate rewards; success = fewer trajectory-level failures
• Increase entropy in synthetic generation: Add controlled randomness to data pipelines; success = measurable diversity in model outputs
• Layer capabilities gradually: Add modalities and memory features step by step; success = stable gains without regression in existing skills
• use autocomplete strategically: Use LLM coding help for boilerplate and scaffolding; success = reduced developer time on repetitive patterns

Conclusion

AI agents will not leap to full autonomy overnight. Building resilient, adaptable systems requires deliberate focus on the cognitive core, careful layering of capabilities, and a willingness to embrace incremental progress. Leaders who align technical design with pragmatic business strategy will be best positioned to harness AI's evolving potential.

For the complete discussion and Karpathy's perspective on AI agents, I highly recommend watching the full interview with Dwarkesh Patel.

Questions or feedback on this analysis? Feel free to reach out!