RefKarpathy: December Coding Agents Breakthrough

Karpathy: December Coding Agents Breakthrough

Karpathy: December Coding Agents Breakthrough

Andrej Karpathy's February 2026 thread documenting the fundamental shift in AI coding agent capabilities that occurred in December 2024/early 2025.

Andrej Karpathy tweet on December coding agents breakthrough

Overview

Karpathy identifies December 2024 as an inflection point where coding agents transitioned from "basically didn't work" to "basically work" - not gradual progress but a sharp discontinuity. The core change: significantly higher quality, long-term coherence, and tenacity enabling agents to power through large, complex tasks autonomously.

Key Example: Home Video Dashboard

Weekend project demonstrating the new capabilities:

Task given (in English):

"Here is the local IP and username/password of my DGX Spark. Log in, set up ssh keys, set up vLLM, download and bench Qwen3-VL, set up a server endpoint to inference videos, a basic web ui dashboard, test everything, set it up with systemd, record memory notes for yourself and write up a markdown report for me"

Result:

  • Agent completed autonomously in ~30 minutes
  • Ran into multiple issues, researched solutions online, resolved them
  • Wrote code, tested, debugged, set up services
  • Returned with complete report
  • Zero manual intervention required

Historical context: This would have been a full weekend project just 3 months prior.

The Paradigm Shift

Old: Direct Code Authoring

  • Typing computer code into an editor
  • The default since computers were invented
  • That era is over

New: Agentic Management

  • Spinning up AI agents
  • Giving them tasks in natural language
  • Managing and reviewing parallel work streams
  • Ascending layers of abstraction

Agentic Engineering

The emerging skill: Setting up long-running orchestrator "Claws" with:

  • Right tools
  • Memory systems
  • Instructions
  • Managing multiple parallel Code instances

Leverage: "Top tier 'agentic engineering' feels very high right now"

Constraints & Limitations

What it needs:

  • High-level direction
  • Judgment and taste
  • Oversight and iteration
  • Hints and ideas

Where it works best:

  • Well-specified tasks
  • Verifiable/testable functionality

The key skill:

Building intuition to decompose tasks appropriately - hand off parts that work, help around the edges.

Implications

Core assessment: "This is nowhere near 'business as usual' time in software"

The shift represents a fundamental transformation in the practice of programming, moving from direct code authorship to agentic orchestration and management.

Context

Karpathy is uniquely positioned to assess this shift:

  • Former Director of AI at Tesla (Autopilot)
  • Founding member of OpenAI
  • Created CS231n (Stanford's foundational deep learning course)
  • Built nanoGPT and other educational AI tools

His assessment carries weight both from technical depth and practical experience building production AI systems.

Commentary

This thread crystalizes a widely-felt but hard-to-articulate shift. The specific December timing aligns with:

  • Claude 3.5 Sonnet improvements (October 2024+)
  • GPT-4 Turbo refinements
  • Emergence of specialized agent frameworks

The "30-minute weekend project" example is particularly striking - not because it's flashy, but because it's mundane infrastructure work requiring:

  • Multiple technology stacks (SSH, vLLM, systemd)
  • Remote system administration
  • Debugging and problem-solving
  • Documentation

The fact that this just works now, reliably enough to "kick off and forget," represents a practical threshold crossing.

Timestamp

Posted: February 22, 2026 (roughly 2 months after the December breakthrough he describes)

Edit this page