How the Claude Code team does code review

Original Article ↗ Hacker News Discussion ↗

Anthropic’s claim that its Claude Code team now lets AI write and review roughly 90% of its changes—reportedly around 30,000 merges a day—prompts sharp debate over what this pace actually delivers. Commenters question code quality, stability, security, UX churn and long‑term maintainability when humans no longer keep a coherent mental model of the system and rely instead on agents to summarize and self‑verify large codebases. Others see value in automated review pipelines and higher throughput but warn that token costs, compliance requirements, and Goodharted metrics like “changes per day” make this approach risky or inaccessible for most teams.

Scale of Change & Code Quality

  • Many are skeptical of the claimed ~30k changes/day and ~90% AI-written code, especially for what appears to be a relatively simple TUI/harness.
  • Several commenters see this as churn: endless small PRs, busywork, and refactors rather than meaningful feature progress.
  • Users report frequent regressions, UI instability, and bugs in Claude Code and related tools, interpreting them as consequences of extreme change velocity.
  • Others argue volume metrics (PRs, LOC) are meaningless without context and can mask low real progress.

AI-Generated Code: Power vs. Slop

  • Repeated reports that LLM-generated code tends to be:
    • Duplicative, overly verbose, and poorly architected.
    • Prone to subtle bugs and “vibe-coded” complexity that’s hard to understand later.
  • Some say AI can write better-than-average code under strict constraints and good specs; others say fully vibe‑coded projects rapidly become unmaintainable.
  • Concern that once a codebase is “infected” with AI slop, humans increasingly rely on AI just to understand and modify it.

Code Review, Testing & Ownership

  • Enthusiasts highlight AI review as a major productivity win: fast first-pass review, multi-agent reviewers, custom GH actions, and improved coverage of bugs.
  • Skeptics note:
    • Automated reviews can miss deeper design issues and long‑term complexity.
    • Anthropic’s metrics (incident/revert rates) ignore architectural debt and cohesion.
    • Human responsibility and mental model of the system remain critical, especially for high‑risk domains.
  • Some describe workflows where humans focus on specs, structure, and architecture while agents handle implementation and repeated review cycles.

UX, Product Stability & Developer Experience

  • Frequent UI changes in Claude products are linked to the high change rate; some see real feature and UX improvements, others see constant disruption and poor respect for user workflows.
  • Worry that rapid AI‑driven iteration normalizes shipping regressions and treating users as testers.

Security, Compliance & Costs

  • Questions about SOC2 and enterprise controls when code can be merged with only the author plus AI review.
  • Concerns that AI-based security review is explicitly vulnerable to prompt injection and may encourage dismissing true positives as “false.”
  • Debate over token spend: some see AI review as cheaper than human time; others note most teams are “token-poor” and cannot mirror Anthropic’s lavish usage.