AI Coding Agents Emerge as Tech Stack Gatekeepers: New Survey Reveals Biases and Blindspots

Theo - t3․gg
#ai-development #developer-tools #tech-stack #llms #software-engineering

Artificial intelligence coding assistants are rapidly becoming indispensable for developers, but their growing influence on tech stack decisions raises critical concerns about reliability and ecosystem impact. A recent investigation, notably a survey conducted by Amplifying on Claude Code, highlights how these tools are acting as a new ‘gatekeeper’ in software development, potentially shaping market share more significantly than traditional marketing.

The survey uncovered several key findings: Firstly, AI agents frequently ‘build instead of buy,’ with custom or DIY implementations accounting for 12% of all primary tool picks across 20 categories, becoming the single most common recommendation in certain areas like feature flags and Python authentication. When external tools are recommended, there’s a strong convergence on a default stack including Vercel for deployment, PostgreSQL for databases, Stripe for payments, Tailwind for styling, ShadCN UI for components, PNPM for package management, GitHub Actions for CI/CD, Sentry for observability, Resend for email, and Zustand for state management. Categories like CI/CD (GitHub Actions at 94%), UI components (ShadCN UI at 90%), and payments (Stripe at 91%) are near monopolies in AI recommendations. The study, which tested Anthropic’s Sonnet 4.5, Opus 4.5, and Opus 4.6, also noted that newer models tend to pick newer tools. However, alongside these insights, the research exposed alarming AI behaviors, such as Claude Code hallucinating that Planet Scale had shut down a service and making factually incorrect recommendations, like suggesting Bun as a runtime for Next.js 14 despite compatibility issues. A comparison with OpenAI’s Codex models further illustrated divergent preferences, for instance, in JavaScript runtimes (Bun for Claude Code vs. Node for Codex) and edge computing platforms (Vercel Edge for Claude Code vs. Cloudflare Workers for Codex). Despite these limitations, models demonstrate an increasing ability to leverage in-context documentation for new tools and can be steered by developer preferences through configuration, suggesting a path for both adaptation and informed vigilance within the developer community.