Early Blog

An impact set is useful only when the team knows how it was produced, what it observed, and what remains uncertain.

I removed one open source system, doubled down on open source, and rebuilt our website as a repository I could control.

Claude Code and Codex made agents capable of doing the work. They did not transfer responsibility for what ships.

The bottleneck in software delivery has shifted from writing code to verifying what changed.

The hard part is no longer spotting ugly code. It is proving a clean change did not alter behavior the model never understood.

A prompt-driven workflow for onboarding a repository, setting up Jest, testing backend and frontend code, and reviewing the results.

A clean diff and passing tests are useful evidence. They do not show every existing behavior a release candidate may have changed.

Choose tools by the bottleneck they remove, the evidence they produce, and the work they add elsewhere.

The category has moved beyond AI-assisted test scripts. The practical question is which kind of evidence your team needs before it ships.

Measure how engineering effort becomes valuable, reliable product change. Code volume, pull request counts, and AI usage do not prove productivity.

The best extension stack is not the longest. Choose one tool for each job and remove anything that does not earn its place.

AI code review checks whether a change is correct. Regression analysis checks what existing behavior the change put at risk.

AI code review earns trust through a clear role, maintained context, bounded permissions, evidence, and accountable approval.

The best coding agent depends on where you work, how much autonomy you want, and what your team needs to control.

Choose a language-native collector first, then add the reporting and governance layer your team actually needs.