Why we're looking at this
We ship websites and full-stack apps as a core part of our business, and the tooling underneath that work has changed faster in the last two years than in the prior decade. We track this closely because it directly changes how we scope and staff client projects — see the broader trend in full-stack trends.
What we're seeing
- Agentic coding tools (working across a whole repository rather than one file) meaningfully shrink the time from a written spec to a working first pass, particularly for CRUD-heavy features and API integrations.
- The bottleneck has shifted from typing code to reviewing it — teams that don't invest in code review discipline for AI-generated changes accumulate subtle bugs faster than they used to.
- Design-to-code handoff is improving quickly, but production-grade accessibility and performance work still needs a human pass; we haven't seen an agent reliably catch these on its own yet.
Open questions we're still chasing
We're actively testing where agent-authored code needs a stricter review bar than human-authored code, and where it's actually held to a higher standard already because it's easier to regenerate than to patch. Related notes: AI code review at scale.