Why The Best Engineers Are Solving Code Review Bottlenecks

TL;DR

• 1

  Code review is now the bottleneck: Florian Buetow says AI makes code generation 10x to 100x faster, which pushes pressure downstream onto review, architecture, and senior engineers' attention.

• 2

  The harness matters more than the model: In Florian's experiments, the same frontier model behaved very differently across Claude Code and Codex, because the harness controls tools, prompting, memory, and execution.

• 3

  Spec-driven development alone was not enough: A "perfect prompt" kept drifting, but pairing specifications with behavioral tests and automated feedback loops finally produced code that matched intent.

• 4

  Guardrails should run close to code generation: Instead of waiting for GitHub PR review, Florian recommends local checks like formatters, linting, Semgrep, security scans, and architecture tests that return natural-language feedback the agent can act on immediately.

• 5

  Architecture is still a human job: He argues engineers should still define what to build and sketch services, modules, interfaces, and constraints up front, because models are not reliable enough to own system design.

• 6

  You can mine session logs for repeated mistakes: Florian suggests analyzing Claude session history to find patterns where you keep correcting the agent, then turning those recurring fixes into static checks or Semgrep rules.