I’ve written before about having an AI employee. Henry runs 24/7, handles code, research, and drafts. That setup with Claude Code has been working well.
But one employee isn’t always enough.
I run StarDust, Aligno, and client work through various agencies, along with a bunch of smaller experiments. Context switching is my default state. And single-threaded AI work doesn’t match multi-threaded life.
OpenAI’s new Codex desktop app might have just solved that problem.
I’m a CLI guy. Have been for years. The idea of a GUI for coding agents felt unnecessary; just give me a terminal.
Then I actually tried it.
The killer feature isn’t the model. It’s seeing all my agents across all my projects in one view. StarDust threads here. Aligno there. Client work below. All running, all visible, all interruptible.
That’s not how Cursor works. In Cursor, I get one window per project. I can’t see everything at once. With Codex Desktop, I can fire off a task, switch projects, start something else, and come back when the first one needs input.
I feel less like a developer now and more like a manager. Checking in on my AI employees, giving direction, moving on. That sounds dystopian but it’s actually… efficient?
I still run Claude Code as my daily driver. It’s fast. Solves problems quickly. Feels good to use.
Codex feels slower. Clunkier at first. But it asks better questions. Way better. The subtle details that Claude misses, and I’d have to catch manually? Codex just asks about them upfront.
I’ve noticed Claude introducing weird bugs that slip through because it moves fast. Codex is more thorough. More careful. For quick wins, Claude. For “please don’t break production,” Codex might actually be better.
So I’m not switching. I’m expanding headcount.
Codex lets you run with full permissions. No approval prompts. It could wipe my home directory and nothing would stop it.
I turned that on immediately.
Reckless? Maybe. But constant interruptions kill flow. I have git. I have backups. The upside is worth it for my workflow.
Know your own tolerance.
While I’m testing Codex, Anthropic dropped a lot yesterday.
Sonnet 5 was supposed to launch but got delayed internally. Word is today’s the day—1 million token context, half the price of Opus, already beating Opus in some coding benchmarks. The leaked demos are ridiculous: fully functional Windows-style web OS in one shot, Celeste clones with real physics.
Xcode 26.3 went agentic. Apple now supports both Claude Agent SDK and Codex natively. The same engine that powers Claude Code is now in the IDE. Model-agnostic with MCP support.
Claude Code v2.1.30/v2.1.31 also shipped. My MacBook finally stopped trying to achieve liftoff—whatever was running it hot seems fixed.
And Claude Code is getting native multi-agent orchestration. Same “manage a team” workflow I’m doing in Codex, but built into Claude.
Right now: Claude for deep work, Codex for parallel tasks across projects.
If Sonnet 5 delivers on the leaks—and with native multi-agent coming to Claude Code—I might consolidate back to one ecosystem.
But honestly? Having options is good. Competition is making all of this better, fast.
Here’s hoping Sonnet 5 actually drops today. My refresh finger is ready.
Links: