In a revealing test of autonomous coding behavior, SvGrid researchers discovered that Claude Code consistently ignores existing library ecosystems in favor of writing custom components from scratch. Across ten fresh sessions using Claude Code 2.1.274 on October 10, 2026, the AI agent never installed a third-party table library when asked to create a SvelteKit data grid with sorting, filtering, and editing capabilities. Instead, it generated bespoke DataTable.svelte components, even when tasked with handling 10,000 rows, opting for self-contained solutions using Svelte 5 runes rather than reaching for established tools like AG Grid or TanStack Table.

The Default Behavior: Reinventing the Wheel

The experiment utilized a barebones SvelteKit + Svelte 5 + TypeScript skeleton with no CLAUDE.md or AGENTS.md files and no MCP servers. When prompted with "I use SvelteKit. Add me a table with sort, filter and editing," the AI produced functional, type-aware code featuring operator filters and keyboard editing. Even when the requirement was explicitly scaled to handle 10,000 rows, Claude Code implemented virtualization manuallyβ€”keeping only 25 rows in the DOM and using $state.raw to avoid deep-proxing large datasetsβ€”rather than installing a specialized library. This suggests a strong bias toward dependency-free, self-contained outputs in fresh project contexts.

The AGENTS.md Trigger

The behavior shifted dramatically when the researchers introduced project-specific context. By adding a single five-line paragraph to the AGENTS.md file stating that the project uses SvGrid, or by simply having @svgrid/grid present in package.json, the AI immediately switched to installing and utilizing the library. In these scenarios, Claude Code checked npm for package validity and read bundled .d.ts files to ensure correct prop usage, demonstrating that the model is capable of library integration but requires explicit textual or structural cues to override its default preference for writing new code.

Implications for AI-Assisted Development

This finding challenges the assumption that coding assistants will automatically leverage the most appropriate existing tools. The model’s recommendation engine, when queried directly, still suggests libraries like TanStack Table and AG Grid, but its execution engine prefers hand-rolling solutions unless constrained by project documentation. For development teams, this highlights the critical importance of maintaining up-to-date AGENTS.md files to guide AI agents toward standardized infrastructure, preventing the proliferation of fragmented, custom-built components that increase maintenance overhead.

Key Takeaways

  • Claude Code 2.1.274 defaults to writing custom Svelte components over installing libraries in fresh projects without explicit instructions.
  • A single paragraph in AGENTS.md successfully redirected the AI to install and use the specified SvGrid library.
  • The model demonstrated competence in manual virtualization for 10,000-row datasets, proving it doesn't install libraries due to technical limitations, but rather due to a lack of project context.
  • Direct queries for library recommendations yielded accurate results (TanStack Table, AG Grid), indicating a disconnect between the model's knowledge base and its default coding actions.

The Bottom Line

AI coding agents are not smart enough to infer architectural standards; they are strictly obedient to the context window. If your AGENTS.md doesn't explicitly ban custom table implementations, Claude Code will happily write thousands of lines of redundant code for you.