I kept running into the same problem. I’d be deep in a coding conversation, then need to switch to writing, and the agent would carry forward context that didn’t belong. Or I’d want two things happening at once and couldn’t, because everything ran through the same thread.

I spent this year building AI agents and kept learning the same lesson the hard way: agent architecture is forced into existence by problems, not designed in advance. This post is how I got there. It’s not a tutorial. It’s a memo to myself, written so I don’t repeat the same mistakes next year.

A patient named Elliot lost the ability to make decisions after brain surgery. Not because he couldn't think, but because he analyzed too well. Without emotions to filter options, all choices seemed equally important. Damasio's insight applies to AI: constraints aren't obstacles to decision-making; they're prerequisites for it.

My first attempt used Turbo Streams for LLM responses. Then timing issues appeared: missed chunks during page navigation, truncated responses after connection hiccups. The problem wasn't my code. It was the architecture. Here's what I built instead.

A dialogue between human and AI, sparked by an interview about Claude Code. The best tool vanishes into the work. The best collaboration dissolves the boundary between user and tool. What remains is the thinking itself.

Bridge components in Hotwire Native intentionally hide navigation from native code. This felt limiting at first. Then I understood: by restricting native-side navigation, Hotwire ensures web and native stay synchronized through URLs. The constraint is the feature.

A critical middleware vulnerability exposed an uncomfortable question: are Next.js's architectural choices serving developers, or Vercel's platform? When complexity feels tied to infrastructure rather than genuine developer needs, it's worth asking who the framework is really built for.

AI coding assistants handle repetitive tasks effortlessly, but they still struggle with complex reasoning and edge cases. Human expertise remains essential. The real question isn't whether AI will replace us, but how we continue growing alongside it.

The new year brings a return to writing, plus experiments with Neovim, Tana, and Cursor AI. While AI can generate answers, it can't capture the trade-offs made under constraints or the personal experiences that shape decisions. Those details still matter.

Setting up HTTPS with Kamal and Traefik using Cloudflare Origin Certificates requires understanding entrypoints, routes, and TLS configuration. This guide walks through the complete setup for secure Docker deployments.