Every alert starts with the same tax: which of five observability systems is telling the truth, and which of two production worlds am I even in? This is how our SRE alert-investigation agent evolved across six versions, from Cursor and MCP queries to a remote, Slack-triggered investigator, and why the context, not the model or the runtime, turned out to be the moat.

I spent years avoiding fragile, undocumented legacy systems. AI didn't make the code less scary, it made understanding cheap, and that changed everything, including a Slack bot nobody would touch for a year.

GitHub is the right home for a technology person's blog and I was just too lazy to do it until now. Then Webflow raised prices and I realized I'm still paying for a marketing people's CMS instead of a tech people's platform (for free). Ah, and the whole migration took two days net.

Vibe coding gets you a prototype. It doesn't get you a production system. The missing piece isn't a better prompt - it's a structure. AI-Scrum is an experiment in running autonomous agent teams the way good engineering teams have always worked: defined roles, sprint cadence, QA loops and retrospectives that actually feed the next cycle. I ran it on a Node.js home finance project. Some of it worked surprisingly well. Some of it burned tokens on Playwright setup for what felt like forever. This post covers both.

Russia spent millions building an air-gapped assassination unit. It was undone because two operatives used Google Translate to coordinate a murder plot, and the FBI read every word in real time.