Karpathy: Coding Agents Have Fundamentally Changed Programming — The Old Era Is Over

A Paradigm Shift in Programming

AI researcher Andrej Karpathy shared an in-depth analysis on X in February 2026 about how AI has transformed programming. He emphasized that this change "is not gradual or in the 'progress as usual' way" — it happened specifically since last December, representing a sharp discontinuity.

Coding Agents Finally Work

Karpathy states that coding agents basically didn't work before December but basically work since. The latest models have significantly higher quality, long-term coherence, and tenacity — capable of powering through large, extended tasks in ways that are extremely disruptive to the default programming workflow.

A Real Example: Building a Home Camera Dashboard

Karpathy shared a concrete example: over the weekend he set up a local video analysis dashboard for his home cameras. He gave the agent a single instruction covering logging into his DGX Spark, setting up SSH keys, installing vLLM, downloading and benchmarking Qwen3-VL, setting up an inference server endpoint, building a web UI dashboard, testing everything, setting up systemd services, and writing a report.

The agent worked independently for about 30 minutes, encountered multiple issues, researched solutions online, wrote code, tested it, debugged it, set up the services, and returned with a completed report. "All of this could easily have been a weekend project just 3 months ago but today it's something you kick off and forget about for 30 minutes," Karpathy noted.

A New Mode of Development

Karpathy declares that the era of typing computer code into an editor — as things have been done since computers were invented — is over. Now developers spin up AI agents, give them tasks in English, and manage and review their work in parallel. He identifies the biggest opportunity as figuring out how to keep ascending layers of abstraction by setting up long-running orchestrator agents with the right tools, memory, and instructions to productively manage multiple parallel code instances.

Caveats

It's not perfect: it still needs high-level direction, judgment, taste, oversight, iteration, and hints. It works better in some scenarios than others, especially for well-specified tasks with verifiable outcomes. But Karpathy is clear: "this is nowhere near 'business as usual' time in software."