Pokémon bakeoff reveals best JavaScript frameworks for AI

JavaScript developers often default to tried-and-true frameworks, but relying on personal preferences can be detrimental when entrusting such tasks to AI agents. Nolan Lawson, an AI security developer, sought to test how various JavaScript frameworks perform when used with AI agents. To do this, Lawson conducted an experiment comparing 12 JavaScript frameworks using the same setup for a basic Pokémon web application.

Lawson evaluated each stack across five independent runs, testing for build cost, maintainability, and end-user performance of the agent-built apps. The most cost-effective build was achieved using Hono paired with htmx, while Astro also performed well with a median cost of $2.76 per Pokéapp. Nuxt was the cheapest option for an SPA-like approach, with the fewest lines of code generated by the AI.

However, more traditional frameworks like Angular and SvelteKit performed significantly worse, with higher costs and longer build times. While the resulting apps all passed accessibility and performance tests, Lawson hopes this experiment will encourage a more rigorous approach to selecting frameworks for AI projects.