Can AI Agents Really Build a Functional Web Browser? Inside Cursor’s GPT‑5.2 Experiment

Experiment Overview

Cursor used hundreds of GPT‑5.2 agents running continuously for about a week to generate a web‑browser prototype from scratch. The agents produced over 1 million lines of Rust code across roughly 1 000 files, targeting a rendering engine, HTML parser, CSS cascade, layout, text shaping, drawing pipeline, and a custom JavaScript virtual machine.

Repository

The generated source is publicly available at https://github.com/wilsonzlin/fastrender. The repository contains the basic browser components described above.

Build Attempts

Both the GitHub Actions CI and local builds using cargo build or cargo check fail. Reviewing the last 100 commits shows no cleanly compilable revision. Example CI error:

错误：由于存在 34 个错误，无法编译“fastrender”（库）；发出 94 个警告

Typical failures include missing module definitions, type mismatches, and unresolved dependencies, resulting in dozens of errors and hundreds of warnings.

Technical Assessment

The codebase lacks coherent architecture; core components such as a DOM implementation or a fully functional JavaScript engine are difficult to locate. The accompanying blog post (dated January 14 2026, URL: https://cursor.com/blog/scaling-agents) provides no reproducible build instructions, known working commit, or runnable demo.

Takeaways

Large‑scale autonomous coding can produce massive code volumes, but without human‑defined design and verification the output may be non‑compilable.

A minimal success criterion for a browser prototype should include successful compilation with a supported toolchain and the ability to render a simple HTML page. The current project does not meet these criteria.

The coordination techniques that allowed hundreds of agents to work on a shared repository may be valuable when integrated into more mature development pipelines.