Blog

Agents can investigate, plan, and implement faster than I can, but reading a final diff is not the same as understanding why the system works the way it does. I own how the system behaves in production, so I want to keep building my intuition about it.

GitHub isn't just falling short on performance and reliability. The whole collaboration paradigm is the wrong shape for how we build software now.

A team process without a principle underneath it is a real problem. Here's why the "why" is worth protecting as you scale.

AI can produce code faster than you can review it, but you're still the one on call when it breaks. You need an AI-assisted workflow you trust: linter guardrails, reference tests the agent can copy, tight AGENTS.md files, and PRs small enough to review. The goal isn't to avoid AI, it's to stay in control of what it ships.

Paranoid hardening has always been with us: aggressive kernel customizations for bug classes without a proven exploit, while Linux itself had a good-enough disclosure track record. Some of these hardenings became the default, some stayed controversial for a long time. AI-assisted vulnerability research made finding obscure vulnerabilities fast and cheap. So what does good enough even mean in this new era?

The longer I worked on a problem with a coding agent, the dumber it felt. My context window was a junk drawer. Here's the multi-role loop I built so plans, builds, and reviews each get their own clean context.

Agents can write and push code in seconds, but CI still takes minutes to give them feedback. That mismatch is the biggest bottleneck in agentic engineering today, and it's why CI needed to be rebuilt from scratch.

Offload the scavenger-hunt grunt work to agents, log the useful bits (also with an agent) so the problem is discoverable next time, and spit out the day's support recap in seconds.

I used an LLM to refactor a 1500-line mess into clean, organized code. I realized months later that agents were the only ones working in it. They didn't need the clean code. They didn't need my empathy. My intuitions as an engineer just don't work anymore.

AI can help write your feature in 20 minutes. But if your CI pipeline also takes 20 minutes, you've just traded one bottleneck for another. The bottleneck in software development has shifted from writing code to integrating it, and our workflows haven't caught up.