GitHub has published a complete retrospective of its 18-month refactor—a rare case of a major company openly admitting its tech path went sideways. Official figures from its design system: server-side rendering time cut by 55%, component initialization reduced by 25%; main site pages improved 1%–22%. The final 3 weeks: 2 engineers paired with Copilot coding agents (GitHub's AI coding assistant) zeroed out the remaining 895 legacy style call sites.
What this is
As of April 2025, GitHub had roughly 7,760 legacy style call sites. The first 6 months: 8 engineers manually migrated 6,419 of them. The remaining 895 went to 2 engineers plus Copilot agents and got cleared in 3 weeks.
The mechanism is straightforward: CSS-in-JS converts style objects into rules and generates class names at browser runtime; traditional CSS is compiled at build time. "Ship a little more CSS, let JavaScript do less"—as component counts rise, the math actually flips in CSS's favor.
Industry view
Supporters call this a real-world case study of AI in production: 8 engineers, 6 months, 6,419 sites vs. 2 engineers + AI, 3 weeks, 895 sites. The value of tools like Copilot, they say, is absorbing the mechanical, repetitive rewrite work.
The dissent deserves more attention. GitHub's self-reported data lacks full experiment windows and page distribution—it's their own result, not a generalizable conclusion. AI only entered in the final 15%—the prior 85% of architectural design, the dual-track compatibility layer (keeping old and new styles coexisting), and the rollback switches were all human-built. For small teams or scenarios where styles change dynamically, migration costs may exceed the gains. And if you're using AI to generate large volumes of new code, you have to bake the new rules into your scaffolding—otherwise AI clears old debt while quietly creating new debt.
Metrics also have traps: watching only JavaScript bundle size shrink while ignoring browser CSS parsing and cache hits; comparing experiment groups across different pages and time windows is another common error.
Impact on regular people
For enterprise IT decision-makers: chasing new tech is not the answer. 90% of GitHub's work was defining performance budgets, designing rollback, and choosing the right migration unit—"boring work." Next time someone pitches a new framework, ask first: can it coexist? Can it roll back? Are there a performance baseline?
For individual careers: AI's real strength isn't writing new features—it's clearing old debts: mechanical rewrites, legacy migrations, doc syncing, code formatting. Hand those to AI; humans should make the calls that need a human.
For the consumer market: sites may be a touch quicker, but users won't notice much. What's worth watching is that counter-intuitive conclusions like "less JS, more CSS" will keep coming—the tech world is rethinking "newer is better."