jsnode

Two years ago, if you told me one person could run an agent workflow that ships app updates while I sleep, I would not believe you. Now I do it as a solo founder.

X search gets sharp once you know what to type, so I put every filter for account, date, likes, and media on one cheat sheet.

Every name for building with AI ends up sounding wrong. "Vibe coding" sounds like there's a candle lit somewhere. "Agentic engineering" sounds like it comes with a lanyard.

Lately I've been trying to point agents at everything: - one to sort my side projects - one to draft YouTube scripts - one to watch my builds while I sleep - one to clean up the mess the other three made Nothing is on autopilot yet. I'm mostly the guy refilling the terminal with coffee and prayers

Nothing says senior engineer energy like an agent marking the task done. The failing test is still there. It's just commented out now, very politely.

People are going to call this a Miro killer. I don’t really care about that part. What gets me is not having to explain my project to Claude every time I open a new chat. My plans are in one place, my numbers are somewhere else, and half my notes are in my head. So I end up pasting the same context over and over before I get anything useful back. If the board and the sheet live right next to Claude, that problem mostly goes away. That’s the part I’m excited to try.

Every productivity app I've tried eventually becomes a second job where I manage the app. The system that's lasted longest is a notebook and a walk when the code stops making sense. The walk fixes more bugs than the terminal does.

Patterns in agents that actually run reliably If you build enough agent workflows, you start to see the same structure every time. Different tools, same architecture. Here's what the reliable ones share: 1. A narrow job They're built for one agent, one task, and one clear output. 2. A safety check before any action They verify first, then move. Automation without guardrails is a gamble. 3. Small, boring steps They're lean on purpose, since every extra step is another place to break. 4. A log you can read When something fails, you can see exactly why, so nobody has to guess. 5. A human in the loop where it matters The decisions that can't be undone stay with a person. The agents I trust are the ones I've watched fail a hundred times and fixed.

Every productivity list tells me to turn off notifications, and then I check my phone to see if anyone replied to the post about turning off notifications. Meanwhile the thing that actually gets me to focus is a deadline and a cold cup of coffee.

"I'm not finished yet." - Stuck (can't ship). - Judged for the empty App Store page. - Has a folder of abandoned projects. "I'm letting it sit for a week." - Patient (could ship but won't). - Respected for taste. - Has a build that is nearly ready.

I used to think every bug was a sign I picked the wrong career. Now I just see a bug and think it's a bug. It's usually a missing comma somewhere.

Build more agents. Building them shows what breaks and where a human is needed. It lets you add safety checks, cut the busywork, and keep your week consistent. It also helps you ship the app and see what Claude Code can do.

Claude Code made me believe in the whole "agents can do real work" thing. A year ago, I didn't think I could run a company solo because there was too much to handle by hand. I would have never gotten here without automation, and now my whole workflow depends on it.

AI agents have finally nailed the coworker who says "quick fix" and then touches 14 files.

Two coding agents on one repo is basically two contractors who never meet. Mine got a lot less chaotic once I gave them a HANDOFF.md at the root. It lists what changed, what's half finished, and what to touch next. Each agent reads it before starting and updates it before stopping. Git log has the what. The handoff file has the reasoning behind it.

knowing how to sell is going to hold up pretty well through all of this. Whatever the tools can do next year, being able to explain why someone should give you their money still seems worth learning.

The developer jobs I can see disappearing are the ones where you're a siloed code writer and not much else. I've met quite a few people like that. They dodged conversations and kept contact with other humans to the bare minimum. AI is coming for those roles first. The engineers who sit between development and other teams or customers are a lot harder to replace.

A few lessons from building AI agents: I started narrow and added tools slowly. I logged everything and trusted nothing. I automated the boring work and reviewed the risky work. Prompts stayed short and permissions stayed shorter. I shipped small and fixed fast.