Three.js is a JavaScript library that can be used to implement 3D graphics in the browser – that is, scenes, cameras, lights and materials, without having to access the graphics card's technical interfaces directly. For a long time, this was a tool for a manageable group of specialists: product configurators, data visualisations, experimental campaign pages. Now the library reports weekly download figures in the double-digit millions.
Such jumps can rarely be explained by organic growth alone. The obvious driver is considered to be AI agents that generate entire web interfaces on request and apparently prefer to reach for a limited set of well-known libraries – Three.js is one of them.
Why models tend towards certain libraries
Language models suggest what is strongly represented and well documented in their training data. Three.js has met both criteria for years: many examples, many tutorials, a stable basic concept that has barely changed over a long period. For a model, this is a safe path – the generated code is highly likely to run on the first attempt.
There is also a reinforcing effect. The more frequently agents use a library, the more code using that library ends up in public repositories, forums and example collections – and thus potentially in future training data. Technology decisions that used to be made in architecture meetings thus become a by-product of statistical frequency.
Download figures are not usage figures
One point of context is important: downloads from a package registry say little about how many production websites a library actually serves. Every automated build, every test environment, every fleetingly generated prototype counts. When agents create, discard and rebuild projects by the minute, high figures arise without corresponding distribution in live operation.
For assessing a technology, this means: download curves are currently even less suitable as a maturity indicator than before. Questions about the release cycle, breaking changes, the size of the maintainer group and suitability for the specific use case remain more relevant.
What this means for performance
3D in the browser is not free. A scene brings additional JavaScript load, textures, geometry data and permanent render loops. This costs loading time, computing power and, on mobile devices, battery. When an agent delivers a homepage with an animated 3D background because that was the common answer to "modern landing page" in its training data, this creates an effort that no one has justified on technical grounds.
Especially in mid-sized companies, this collides with hard targets: Core Web Vitals as metrics for loading behaviour and interactivity, accessibility requirements, energy and hosting costs. Moving 3D elements can also exclude users who depend on reduced animations – a point that generated code rarely takes into account on its own.
Maintainability: the real cost factor
A 3D component brings its own terminology. Anyone who is to change it later needs knowledge of scene graphs, materials and render pipelines – knowledge that is not available in many teams. Code that is created in seconds may have to be maintained for years.
The pattern is not new, but agents are giving it a new speed. Dependencies grow faster than review processes can absorb them. What matters is therefore less whether AI-generated code is used, but whether it goes through the same review steps as hand-written code.
Practical guardrails
- Approve dependencies, don't discover them afterwards: New libraries in generated code belong in an explicit review, including licence, maintenance status and bundle size.
- Define budgets: Fixed limits for JavaScript size and loading time make it visible when a generated component exceeds the framework.
- Put specifications into the prompts: Anyone who provides their own technology stack, permitted packages and accessibility rules as context significantly reduces random selection.
- Separate prototype and production: Agents are strong at rapid illustration. For live operation, a conscious decision is needed on whether the prototype stack is adopted.
- Check appropriateness: A product configurator justifies 3D. A decorative background animation on a corporate website rarely does.
Assessment
The Three.js case shows less a problem of the library than a shift in the decision chain: technology selection is moving from the architecture discussion into the training data of generative tools. For web and software projects, this means actively organising governance over dependencies instead of leaving them to a model's suggestion behaviour. Anyone who establishes performance budgets, approval processes and clear stack specifications uses the speed of AI agents without buying technical debt that only becomes visible in the maintenance phase.
Sources
- Hype around Three.js: Why AI agents love this JavaScript library — t3n.de - News