Framework debates used to be about developer experience, bundle size, and maintainability. Nolan Lawson has added a new variable: how much does it cost an AI agent to build something in this stack.
The methodology is concrete. He asked Claude Opus 5.5 to build a Pokémon app to a fixed spec in roughly a dozen frameworks — React, Vue, Svelte, SvelteKit, Astro, Hono+htmx, Next.js, Nuxt, Angular, Lit, Gracile+Lit, and Vanilla Vite — running each configuration three to five times to account for variance. He measured cost, wall-clock time, token usage, Lighthouse scores, and lines of code produced.
The spread on cost was almost 2x: Hono+htmx came in at a median of $2.67 per run, Angular at $4.97. Astro was the fastest (8.2 minutes), Angular the slowest (17.2 minutes). On generated JavaScript payload, the range was enormous — Astro delivered near zero, Next.js produced about 142 KiB. Multi-page frameworks generally won on Largest Contentful Paint.
The likely mechanism isn’t surprising once you name it: verbose frameworks with more ceremony — more boilerplate, more configuration surface, more file scaffolding — give the agent more work to do and more opportunities to go down wrong paths. Angular’s TypeScript ceremony and file proliferation create more tokens in and more tokens out, even on a small app. Hono+htmx stays terse.
What makes the post interesting is its explicit empiricism. Lawson’s point isn’t “use Hono+htmx.” He’s asking whether teams making stack decisions for agentic workflows are thinking about this at all, or whether they’re still choosing frameworks based on community familiarity and hiring pools while ignoring a cost variable that can double their inference budget. For a team doing heavy agent-assisted development, a 2x cost difference compounds over millions of agent runs.
The caveats are real: one small, content-focused app is a single data point, agent behavior was noisy, Angular builds sometimes failed the spec outright, and the conclusion might not hold for complex SPAs with real data-fetching requirements. Still, it’s more evidence than most teams have.
A different angle on AI changing software economics appeared the same week, from an unexpected corner of the open-source world. PostgreSQL committer Robert Haas noted that Postgres 19 is delayed because AI tools are finding latent bugs in the codebase faster than the team can fix them. Back-patched commits have surged. Several features were reverted late in the beta cycle. The project is healthy — just overwhelmed.
The contrast with the framework cost story is worth noting. In one case, AI is a cost to be managed: a tool that consumes more resources when it faces more complexity. In the other, AI is a forcing function: a tool powerful enough to surface decades of accumulated technical debt in a mature, carefully maintained codebase.
Neither observation is about AI replacing engineers. Both are about AI changing the shape of the work — what it costs, what gets surfaced, what blocks a release. The teams adapting are the ones measuring.