Agentic Coding: Bet on the Primitives

A freelance project recently needed three charts — a stacked bar, a line, and a donut — and they had to match an existing design system exactly. The usual answer was already in place: Recharts was installed, and for years that would have been the obvious move. But this time the author tried something different, because agentic coding changed the cost math enough to make the experiment possible.

Within an hour, two versions existed side by side. One used D3’s math primitives with React handling the SVG. The other used Recharts with the same component API and the same design tokens, so the swap was supposed to be clean. The primitive version fit the design precisely, including entrance animations, crosshair tooltips, and legend filtering. The Recharts version got most of the way there faster, then bogged down in custom shapes, custom tooltips, custom labels, and finally a frozen entrance animation that had to be turned off.

That result points at the real issue: high-level libraries are prepaid labor, not free labor. They save time because somebody else already spent the hours figuring out axis ticks, tooltip behavior, legend toggles, and all the annoying edge cases. But that bargain comes with opinions. Once your needs drift off the happy path, you start paying again in wrappers, patches, escape hatches, and the sort of GitHub issue that ends with “not planned.”

The piece places this in a longer web history. jQuery plugins, Backbone, Angular, React, component kits, ORMs, create-react-app, and later meta frameworks all moved work out of your codebase and into dependencies. That was rational when human implementation time was the scarce thing. The new claim is that agentic coding has made implementation labor cheap enough that the pendulum can swing back toward primitives plus a thin layer your team actually owns.

That does not mean every dependency is bad. The author is explicit that d3-scale, date-fns, and zod still make sense because they compute things for you without deciding your product for you. The bet is narrower: own the rendering, interaction, and behavior that give the product its identity, and keep the truly hard, stable problems in libraries. In the chart example, that meant a few hundred lines of custom code instead of a framework that almost fit and then refused to bend.

There are limits. If a team lacks an agentic workflow, the old math still applies. If the UI is commodity, a mature library still wins. And for things like comboboxes, date pickers, and modals with tricky focus handling, the warning label should stay on the box. But for custom flows that once lived too far from a library’s happy path, the crossover point has moved. The article’s real argument is simple: when code is cheap to write, it becomes expensive not to own.