Latest from the Blog

Listen to the error messages in your team!

Some engineers still reject AI agents in their daily work. Why is that?

DORA metrics reveal throughput and stability of your delivery — but not what happens inside. The SPACE framework adds five important dimensions to the picture, and in this final part I show you which three are actually worth measuring once AI agents join your team.

AI agents like Codex somehow aren't making your team faster? Maybe this is why:

Part one covered how not to measure your team's performance. Now for the part that actually helps: the five DORA metrics, what each one tells you, how to measure and automate them — and why, in the age of AI coding agents, they matter more than ever.

Token leaderboards, lines of code, and Goodhart's Law: the metrics most teams reach for to judge AI's impact measure the wrong thing — and actively work against the very efficiency they promise. Here are the three mistakes I run into most, and a first look at what actually works.

AI agents reliably raise individual engineer output. But real gains only land inside well-run organizations — everywhere else, the higher output simply exposes where things really get stuck.

Claude Code plugins cannot all be treated the same. Some can safely be checked into the repo; others are better left out — and the difference lies in what they do to the way people work.

Here are the AI experiments every one of you should actually be running right now; especially if you build software.

Managers and engineering leads keep asking me: how much faster do we really get when we use AI agents in software development? Here is what five major studies — Faros AI, METR, Google DORA, GitLab, and Capgemini — actually reveal once you put telemetry data and self-reports side by side.

Here are three things that are guaranteed to go wrong if you introduce AI agents into your software development right now:

If I could change one thing about Claude Code right now, it would be how it handles memories. I'm firmly convinced that the memory system makes Claude Code worse, not better.

If anything positive comes out of this whole AI craze, then please let it be that we have finally learned how to describe requirements clearly.

A few weeks ago, I was still certain that our tech jobs were safe. They probably still are, but they will change fundamentally.

If you've developed mobile apps for more than a few weeks, you know the pain. You start a project, build a few features, and run it. After a few weeks, building manually becomes tedious, so you check the only real option to automate the process: Fastlane. And while the initial setup seems straightforward, chances are your terminal will still explode with red text.

Everyone tells you to collect feedback fast. Everyone tells you that less is more. Everyone tells you the cheapest feature is the one you don't need and therefore don't have to build. And you keep hearing that at the beginning you're allowed to be “quick and dirty” to get your software idea to market fast.

Over our first year in Bangkok, a habit has formed: at dinner we often watch an episode of Gordon Ramsay's "Kitchen Nightmares." Ramsay's rough manner and creative insults lighten the mood after a long workday, even if what he sometimes finds in restaurant kitchens can definitely kill an appetite.