Linus Torvalds vibe-coded a Python audio visualizer in January

Why it matters

Torvalds' experiment shows the credible near-term use of coding agents: delegate a bounded, low-risk task, then diagnose, redirect and review the output before merging it.

Linus Torvalds used Google's Antigravity coding agent to rewrite a Python visualizer for AudioNoise, his personal digital-audio project, according to a GitHub commit from January 7th, 2026.

The experiment resurfaced Friday after yash.jsx (@yashmp2004) posted a screenshot describing it as confirmation that the Linux creator had embraced vibe coding. The underlying work is real. The timing is seven months old, and the scope is considerably narrower than the viral framing suggests.

The commit in Torvalds' AudioNoise repository merged a branch called "antigravity" and changed one file, visualize.py, with 331 additions and 132 deletions. The listed changes added custom zoom selection, keyboard navigation, high-zoom sample markers, layout revisions and memory-use improvements.

Torvalds described the result with characteristic economy: "Is this much better than I could do by hand? Sure is."

A coding agent with supervision

The commit also records where Google's agent went wrong. Antigravity initially used a built-in rectangle-selection component that caused trouble. Torvalds identified the source of the problem and instructed the agent to implement a custom RectangleSelector. The work then proceeded more successfully, he wrote.

That sequence matters. Antigravity generated a substantial revision, but Torvalds still had to diagnose the failure, change the technical direction and review the resulting patch. The human contribution moved from typing each line to specifying and validating the implementation.

The repository's README is even more explicit about the division of labor. Torvalds says the Python visualizer was "basically written by vibe-coding" because he knows less about Python than he does about analog filters. He had started by searching Google and adapting examples before using Antigravity to remove himself from much of the implementation loop.

AudioNoise is a learning project built around simulations of digital guitar effects. Torvalds describes its filters and delay loops as basic experiments, with single-sample input and output rather than the neural-network cabinet emulation found in some commercial guitar equipment. The commit belongs to that hobby repository, not the Linux kernel.

That boundary makes the work more instructive than a blanket claim that one of software's most consequential engineers has handed programming to an AI. Torvalds chose a non-critical visualization layer written in a language he does not consider his specialty. He retained enough understanding of the desired behavior to catch a bad implementation choice and redirect the agent.

An early Antigravity use case

Google had introduced Antigravity on November 18th, 2025, less than two months before Torvalds' commit. Google positioned the product as an agentic development environment able to plan and execute work across an editor, terminal and browser, rather than limiting it to line-by-line code completion.

The AudioNoise patch provided a compact example of that pitch. Antigravity handled a cluster of connected interface changes inside an existing tool, including navigation, zoom behavior and memory management. It also exposed the practical limit hidden by the product category's autonomy language: the agent needed a developer who could recognize why the first approach was failing.

Torvalds' choice of task was equally revealing. The Python script helps him inspect audio samples produced by code elsewhere in the project. A defect could make the visualization confusing or inaccurate, but it would not enter a production operating-system code path. That gave him room to trade direct control for speed while learning a new technical area.

The renewed attention on August 28th turns an old commit into a fresh endorsement story. The repository supports a narrower conclusion. Torvalds used an AI coding agent, found it productive for a personal Python tool and documented both its useful output and its need for human correction. That is supervised engineering work, even when the machine writes most of the diff.