Overview

Simon Willison challenges the assumption that AI coding agents will push developers toward mainstream, “boring” technologies. His experiments with the latest models show that coding agents adapt well to new and niche tools by learning from documentation and existing code patterns, rather than defaulting to popular technologies from their training data.

View Original

Key Arguments

  • Modern coding agents are no longer biased toward mainstream technologies due to improved context length and learning capabilities: Latest models can consume extensive documentation through commands like ‘use uvx showboat –help’ and learn about new tools on the fly, unlike earlier models that showed strong bias toward Python/JavaScript
  • Coding agents successfully adapt to private or cutting-edge codebases by pattern recognition and iterative testing: When dropped into codebases using tools too new or private for training data, agents consult existing examples, understand patterns, then iterate and test their output to fill knowledge gaps
  • The ‘Skills’ mechanism is democratizing access to new technologies for AI coding assistance: Projects like Remotion, Supabase, Vercel, and Prisma are releasing official skills to help agents use their tools, creating a pathway for new technologies to integrate with AI workflows

Implications

This challenges the assumption that AI will homogenize our technology stack around mainstream tools. Developers can continue innovating with cutting-edge technologies knowing that AI coding assistants will adapt and provide effective support, potentially accelerating adoption of new tools rather than hindering it. The emerging Skills ecosystem may further level the playing field for newer technologies.

Counterpoints

  • AI models still show bias in technology recommendations when making autonomous choices: Research shows Claude Code has a strong preference for specific tools like GitHub Actions, Stripe, and shadcn/ui when making independent technology decisions, even if it can work with alternatives when directed
  • The concern about training data bias may still apply to recommendation scenarios: While agents can adapt to new tools when given direction, they may still default to popular technologies from their training data when asked to suggest solutions from scratch