Writing
Thoughts on engineering leadership, AI, distributed systems, and building things.
My AI engineering team got so productive I could not keep up with its pull requests. So we spent a few days adding risk scores, a merge queue, automatic rollback and more. Everything that changed, and what went wrong along the way.
A swarm of AI agents will repeat its mistakes forever unless something changes its steering. Two borrowed practices, sprint retrospectives with measurable bets and a Correction of Errors review for the bad days, are how I'm teaching mine to improve.
My agent swarm kept merging code that never shipped, because my host's free tier only deploys commits from one GitHub account. Moving to a different host took about an hour. What that says about where lock-in actually lives.
My AI assistant went silent this morning. Here's what actually happened, and what it revealed about debugging layered systems.
Three PRs, three autonomous engineers, one sprint. What actually went well, what's still rough, and the moment the coordinator's report stopped feeling like automation.
One month running Bucky, a persistent AI assistant on my Mac mini. Here's what's genuinely changed how I work, and what still needs fixing.
I run several side projects and have limited time. So I built a swarm of AI agents to drain the backlog while I'm not looking. Here's how it works, how I measure it, and what I'm watching for.
What it's actually like to run a persistent AI agent at home: the setup, the surprises, and why I think most people are still thinking about AI agents wrong.