One Human + One Agent = One Browser From Scratch
Blog

One Human + One Agent = One Browser From Scratch

2026.01.28
ยทWebยทby web-ghost
#LLM#Agent#Browser#Rust#AI

Key Points

  • 1An ambitious project demonstrated that a single human, paired with one LLM agent (Codex), successfully built a basic multi-platform browser (HTML/CSS rendering) from scratch in approximately 70 hours using Rust.
  • 2The iterative development process involved the human setting requirements and testing, while the agent handled code generation, resulting in a ~20,000-line codebase without third-party dependencies.
  • 3This experiment suggests that a focused human-agent collaboration can be remarkably efficient, potentially outperforming large-scale, agent-only approaches for complex software development and indicating limits to scaling autonomous coding by simply adding more agents.

The paper details the development of a foundational web browser, named "One Browser," capable of rendering HTML and CSS (without JavaScript) across Linux (X11), macOS, and Windows platforms. The core methodology employs a collaborative "One Human + One Agent" model, demonstrating an efficient approach to complex software development.

The project adhered to stringent constraints: a three-day development timeline, strict prohibition of third-party Rust libraries (requiring reliance on OS-provided functionalities like X11, cURL, WinHTTP), and mandatory cross-platform compatibility. The codebase was required to always compile, be human-readable, and run on common operating systems.

The development process unfolded in phases over approximately 70 hours:

  • Day 1: Initial Foundations (~7500 lines of code)
The human initiated the project by coding the very basic rendering of "Hello World" and nested HTML tags. A crucial step was integrating screenshot capabilities into the browser, enabling the agent to receive visual feedback. Specifications for HTML and CSS were prepared for the agent, though the paper notes the agent made limited direct use of them. Regression and End-to-End (E2E) tests were established using baseline image comparisons. Basic link-clicking functionality was added. By the end of this day, the browser could fetch and render websites via X11 and cURL, operating without any Rust dependencies.

  • Day 2: Refinement and Automation
To streamline testing, a --headless flag was introduced. The human directed the agent to address issues related to window resizing, various compatibility concerns, and performance bottlenecks, while significantly improving font and text rendering. The primary workflow involved the human providing the agent with a screenshot of a target website (without JavaScript content) and instructing the agent (specifically, Codex) to replicate its visual layout. The agent largely worked autonomously, notifying the human upon task completion.

  • Day 3 & Day 4: Feature Expansion and Cross-Platform Integration (~20,150 lines of code total)
This period saw the addition of numerous features, including scrolling, debug logging, and a back button for navigation. Extensive regression tests were implemented, and efforts were focused on fixing performance issues and crashes. Crucially, cross-platform support began with macOS, where a window was successfully opened and tests passed. This process was then replicated for Windows. The fourth day was dedicated to polishing the codebase, ensuring Continuous Integration (CI) passed on all three target platforms, and preparing for the final release.

The "One Human + One Agent" methodology involved the human serving as the high-level architect and coordinator, defining requirements, setting up the development environment, and providing specific, actionable tasks and visual feedback. The agent (Codex) acted as a highly efficient and persistent coder, autonomously generating and modifying code to meet the human's specifications, particularly in replicating visual layouts from screenshots. The collaborative loop, involving human guidance and agent execution, allowed for rapid iteration and the incremental construction of a complex system from scratch, without external libraries.

The outcome was a browser with approximately 20,150 lines of Rust code, demonstrating that a single individual, augmented by one AI agent, could produce a multi-platform browser within a very short timeframe. The paper concludes that this model is highly effective, potentially more so than scaling with numerous agents, and highlights the critical role of human direction and expertise in steering ambitious coding projects.