Fluent Code, No Taste

The number that stopped me this week: financial services teams have gone from hundreds to thousands of lines of code per day with tools like Cursor — and there aren't enough engineers to review what's coming out. Generation has lapped comprehension, and the gap is compounding.

Luca Mezzalira put his finger on why this is worse than a staffing problem. AI agents produce fluent, functional code at scale, but they lack the judgment experienced engineers build over years of maintaining systems: the instinct that tells a clean design from a fragile one, an abstraction earning its complexity from one adding it. Without that judgment, his phrase is exact — agents become sophisticated amplifiers of mediocrity. The architectural equivalent of a grammatically correct essay that says nothing.

There's a bias underneath that I think about a lot. The training corpus skews toward complex, impressive-looking solutions, because that's what gets written up. Yusuf Aytas's piece on over-engineering describes humans reaching for Kubernetes when a monolith would ship faster; agents trained on the internet inherit that same reflex, amplified. Left alone, they will confidently build you a distributed system you didn't need.

So what do you actually do? Two things, I think. First, redesign review. Our review processes assume humans are the bottleneck on generation; that world is over. Mechanical inspection has to move into automated gates — linting, architectural checks, quality thresholds — so the humans you have spend their attention on judgment calls, not formatting. Second, encode taste into the toolchain itself. An early open-source attempt called skills-forge packages clean-architecture principles and quality gates for coding agents. It's young, but the direction is right: if the agent doesn't know what good looks like, the constraints around it have to.

The rest of the week's news reads like context for the same shift. Cursor 3 was rebuilt from scratch as an agent console — the developer as supervisor and systems thinker rather than keystroke craftsperson. Apple, Google, Microsoft, and Anthropic pooled resources on Project Glasswing, hunting vulnerabilities in shared critical infrastructure with an unreleased Anthropic model; when fierce competitors cooperate, they've all smelled the same risk, and I'd bet it's the code flood. Meanwhile Aspire 13.2 shipped a TypeScript AppHost in preview, which pushes TypeScript further into infrastructure territory, and running models like Gemma 4 locally with Claude Code compatibility became genuinely practical for privacy-constrained teams.

The counterintuitive summary: as generation gets cheap, craft gets more valuable, not less. Code volume stopped being the constraint. Knowing what good looks like — and enforcing it — is the work now.