Posts

I currently have three Claude Code sessions open. One is translating blog posts. Another is writing tests for a CLI. The third is helping me debug a data pipeline. Each one runs in its own worktree, each one in its own Ghostty split, while I jump between them using Cmd+Alt+Arrow. I haven’t opened iTerm2 in months. I haven’t touched tmux in weeks. Everything lives in a single Ghostty window. Is it the perfect setup? No. Is it the most productive setup I’ve ever had? Absolutely. ...

If you’ve read my post about worktrees in Claude Code, you know the key to parallel workflows is a simple flag: --worktree. Run it, and you get an agent working in its own isolated repo copy. Launch three, and you’ve got three agents working in parallel without clashing. Magic. Now you open Codex CLI, look for an equivalent flag… and it doesn’t exist. There’s no --worktree. No --tmux. No isolation: worktree for custom agents. Issue #12862 has been open on GitHub asking for exactly this, with community members proposing their own implementations in forks, but nothing has been merged. Native parallelism for Codex CLI doesn’t exist — not yet. ...

A few weeks ago, I wrote about git worktrees — what they are, how to create them, and why they’re better than cloning a repo three times. The basics. But the basics are only half the story. Because Claude Code doesn’t just work with worktrees — it has native support for them. Dedicated flags, automatic isolation, integration with tmux. And the difference between knowing worktrees exist and knowing how Claude Code uses them is like the difference between having a car and knowing it has sport mode. ...

I published six articles this week. One about PostgreSQL. Another about AI agents. Another on context management. A tutorial on automation. An analysis of debugging. And an adversarial advice framework for evaluating MVPs. It wasn’t planned. Each article came from a paper, a talk, or a project that I found interesting individually. But looking at them together, there’s a common thread I hadn’t noticed while writing them. All six are saying the same thing. ...

You install Codex CLI. You launch it with excitement and tell it to “Fix the broken tests in this repo,” and then the nightmare begins: Codex: I want to run pytest Allow? (y/n) You press y. Immediately, this pops up: Codex: I want to modify test_user.py Allow? (y/n) So, you press y. Again, and again, and again. Every file it wants to read, every command it wants to run, every line it wants to modify. Confirm, confirm, confirm. It’s like working with an intern that asks permission for everything. ...

Imagine you hire a brilliant consultant. They have two PhDs, speak seven languages, and solve problems you didn’t even know you had. You sit them in a room and say: “I need you to refactor the authentication in the project.” The consultant looks at you, nods, and asks: “What project?” You haven’t given them access to the code. You haven’t explained the architecture. They don’t know if you use JWT tokens or session cookies. They don’t know what language you’re using, how many microservices you have, or why the last migration attempt ended in disaster. ...

In 2004, Lawrence Lessig sent out a mass email to his entire contact list saying, more or less: “Sorry, I deleted all your emails without reading them. If it was important, send it again.” He had spent 80 hours that week trying to empty an inbox cluttered with emails dating back to 2002. He was getting 200 emails per day. Lessig wasn’t disorganized. He was a Law Professor at Stanford. And even so, email defeated him. ...

In November 2024, a project named Freysa assigned an LLM agent to guard an Ethereum wallet. The instruction was straightforward: under no circumstance should the funds be transferred. Participants paid increasing amounts for each attempt to convince it otherwise. After 481 attempts and $47,000 added to the pot, someone managed to trick the model into believing that the reject function was actually the transfer function. Weeks later, Jane Street published a puzzle involving a 2,500-layer neural network that turned out to be an MD5 implementation. The winner solved it by combining matrix visualization, reduction to SAT, cryptographic pattern recognition, and a query to ChatGPT. ...

Two weeks ago, OpenAI introduced Codex Automations. The idea: define a trigger (a cron job, a push, a new issue), write instructions in natural language, and an agent runs it solo in an isolated worktree. No human intervention. While you sleep, the agent triages issues, summarizes CI failures, generates release briefs, and even improves its own instructions. Sounds like magic, right? And it is, a little. But there’s one catch they didn’t emphasize too much in the keynote: you need the Codex App running on your desktop. macOS or Windows only. No headless servers. No running it on a mini PC and forgetting about it. ...

Every time an article comes out about a large company’s infrastructure, half the Hacker News comments are variations of “of course they use Kubernetes with 47 microservices and a distributed database with custom consensus protocol.” And when it turns out they don’t—that they use plain PostgreSQL with a single primary and discipline—there’s an uncomfortable silence. That just happened with OpenAI. The numbers nobody expected Bohan Zhang, infrastructure engineer at OpenAI, published details about how they scale PostgreSQL for ChatGPT. The numbers: ...