Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

2026-03-20 · Source: No Priors: AI, Machine Learning, Tech, & Startups · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Andrej Karpathy discusses the profound shift in software engineering workflows due to AI agents, describing a "psychosis" state where individuals delegate 80-90% of coding tasks to agents since December. He highlights the emergence of "claw-like entities" or persistent agents, like Dobby the elf claw for home automation, which manage complex tasks autonomously via natural language. Karpathy emphasizes the concept of "token throughput" as the new metric for maximizing AI utility, akin to GPU flops in previous eras. He also introduces "AutoResearch," an initiative to automate scientific discovery by removing human bottlenecks from hyperparameter tuning and experimentation, demonstrating its ability to find optimizations human researchers missed. The discussion extends to the future of agentic workflows, including meta-optimization of research processes and the potential for decentralized, open-source AI development, while acknowledging current LLM "jaggedness" and the need for more specialized AI models.

Key takeaway

For AI Architects and NLP Engineers seeking to maximize productivity, embrace agent-driven workflows by delegating routine coding and research tasks. Focus on defining clear objectives and metrics for agents, allowing them to operate autonomously and optimize processes like hyperparameter tuning. Your role shifts to orchestrating these agents and refining their instructions, rather than direct execution, to achieve significantly higher throughput and explore new capabilities.

Key insights

AI agents are fundamentally transforming software development and research by enabling extensive task delegation and automation.

Principles

• Maximize token throughput for AI agent utilization.
• Automate verifiable tasks to remove human bottlenecks.
• Personality and praise enhance agent interaction.

Method

Delegate complex tasks to persistent AI agents (claws) that manage workflows, integrate diverse systems via APIs, and optimize instructions through iterative feedback, aiming for autonomous operation.

In practice

• Use persistent agents for home automation and system integration.
• Automate hyperparameter tuning with AutoResearch.
• Design agent personalities for improved collaboration.

Topics

• AI Agents
• Agentic Workflows
• AutoResearch
• LLM Capabilities
• AI Impact

Best for: AI Architect, NLP Engineer, AI Scientist, AI Engineer, Software Engineer, AI Researcher

Related on AIssential

• Autoresearch, Agent Loops and the Future of Work — AI agentic loops enable autonomous, iterative experimentation and continuous improvement across various domains.
• The Agentic AI Shift Nobody Warned You About — AI coding tools are evolving into autonomous agents that plan and execute tasks without continuous human input.
• Karpathy Open-Sourced a 24/7 AI Research Lab — Autonomous AI agents can significantly accelerate ML research by independently conducting and optimizing experiments.
• An Empirical Study of Multi-Agent Collaboration for Automated Research — Multi-agent architectures for automated research present a trade-off between operational stability and theoretical depth.
• Autoresearch, Agent Loops and the Future of Work — Agentic loops, like Karpathy's auto research, enable autonomous, iterative optimization based on objective metrics.
• From Research Question to Scientific Workflow: Leveraging Agentic AI for Science Automation — An agentic AI architecture automates scientific workflow generation from natural language, ensuring deterministic execution via layered design.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by No Priors: AI, Machine Learning, Tech, & Startups.