Should AI coding agents help us to remove things?

I had a React 18 project. Needed to migrate it to React 19. The crazy part about 2026 — I couldn't bring myself to do the migration by hand; that would be an insane thing to do. So of course I asked Kiro to convert React 18 → 19. Of course the server didn't start after the conversion. Of course the agent fixed it. But watching Kiro patch up my shiny new React 19 code, I started seeing lots and lots of problems with React's architecture — state management, routing, hooks, refs… huge constructions built to solve things that were never solved in the framework's fundamentals. Okay, a few hours later the project was migrated, up and running.

Well, that was quick. Why not ask the agent to convert React 18 to Svelte while we're at it? Who cares about frameworks anymore! And of course the agent did it. And of course the server didn't start. And of course I asked the agent to fix it — and to use Playwright to make sure the new and old codebases produced pixel-by-pixel identical results on screen.

Once all three codebases were up and running, I asked the agent to compare them. Here's the result - see the picture in the header of the post.

My thoughts looking at the comparison:

• Okay, who cares
• Well… 3× fewer tokens — I can afford a bigger model now.
• Less code → fewer bugs → easier (and faster!) for the agent to reason about.

Then I asked the agent, "Why is that?!" Here's what it said:

• React pulls in 64 runtime dependencies. Svelte pulls in 4.
• Svelte doesn't need a state management library — $state / $derived / $effect are built in.
• Svelte doesn't need CSS-in-JS — inline styles and <style> blocks just work natively.
• React Context → a module-level $state singleton (simpler, no Provider wrapper).
• ReactDOM.createPortal → just {#if} with position: fixed (no separate API needed).

(please note that those agent's comments are valid solely for my code base; please note that 'my code base' means not yours code base)

Okay, now we're actually talking architecture.

And here's the discussion playing out in my head right now — two completely contradictory vectors for the future:

Option A — foundations don't matter anymore. Who cares how clean the code is? An agent writes it, an agent reads it, no human ever opens the file. Pull in 300 dependencies, stack abstractions on abstractions — the agent will deal with it. DX is irrelevant, because a developer isn't a text-typer anymore. DX is the AI's problem now.

Option B — foundations matter more than ever. Agents are the first tool powerful enough to strip away the layers of patches, workarounds and community duct tape we originally needed only because the foundation wasn't strong enough. They can expose the underappreciated cores underneath and work with them directly. The historical advantage of a "huge ecosystem" was really "a huge pile of patches compensating for missing pieces" — and an agent can generate those missing pieces on demand, cleanly, against a solid base.

Are we at the point when AI agents would REMOVE things, and not only create new ones?

I have no idea what the future looks like — but I can't wait to see it. Buckle up.