Agentic OS & App
Bridging the Chasm: Practical AI Experiments in the Servo Engine
Can AI-assisted coding significantly accelerate the implementation of complex web standards? This talk explores a series of experiments on a fork of the Servo web engine, aimed at closing the feature gap with dominant engines.
Experiment 1: Architect as Director (WebNN)
How does a developer with deep architectural knowledge of a system—but little domain knowledge of the specific feature—use AI? In this case, we treat the AI as a high-speed "typist" and implementation partner. We explore using AI to bridge the gap between high-level structural intent and the low-level boilerplate required for a new API, while the human maintains the "integrity" of the system’s core.
Experiment 2: Novice using Test-Driven AI (CSS Grid)
What happens when the human knows nothing about the domain? This experiment focuses on implementing the complex CSS Grid specification by delegating the logic to AI, using the WPT (Web Platform Tests) and the standard as an absolute "ground truth." We analyze whether a rigorous test suite is sufficient to steer an AI through a black-box implementation, and where this "test-driven" autonomy eventually hits a wall.
Experiment 1: Architect as Director (WebNN)
How does a developer with deep architectural knowledge of a system—but little domain knowledge of the specific feature—use AI? In this case, we treat the AI as a high-speed "typist" and implementation partner. We explore using AI to bridge the gap between high-level structural intent and the low-level boilerplate required for a new API, while the human maintains the "integrity" of the system’s core.
Experiment 2: Novice using Test-Driven AI (CSS Grid)
What happens when the human knows nothing about the domain? This experiment focuses on implementing the complex CSS Grid specification by delegating the logic to AI, using the WPT (Web Platform Tests) and the standard as an absolute "ground truth." We analyze whether a rigorous test suite is sufficient to steer an AI through a black-box implementation, and where this "test-driven" autonomy eventually hits a wall.
