Notes on AI coding assistants, early September 2025

I let Claude Code do nearly all of the terraform work of setting up AWS infrastructure and just supervised. The experience gave me a taste of agentic ops and a rather fresh shock. The fluency with which it handled terraform and aws commands was striking. Compared to a general programming language, the more limited rules, narrower domain, clearer scenarios, and well-organized docs and usage examples probably worked in its favor.

It's been barely half a year since I started using Claude Code. The feel of today's coding agents is completely different from just a few months ago. Around last week, Threads suddenly overflowed with posts and comments praising Codex. Right up to leaving my job I'd been using Cursor with gpt-5 happily enough, but agentic coding meant Claude to me, so I didn't even glance at Codex. I had actually installed it once just before leaving, thought it looked pretty rough, and felt vindicated. And yet, sitting down to use Codex properly this time, it felt cleaner than Claude Code.

Claude Code felt like coaxing a rambunctious junior who was itching to code. The code it wrote felt abstracted at a rather low level, and messy.

Codex, rearmed with gpt-5 and with reasoning effort set to high, delivered an entirely new experience. It behaved like a mid-level engineer with good judgment: toss over a reasonable set of requirements and it returns a fairly convincing result. The code it wrote had a stiffer, more conservative feel. If I'd been building something flashy it might have read as boring code, but when you need production-grade code right now, nothing is more reassuring.

Codex's weaknesses: overwhelmingly slow speed, and CLI features less mature than Claude Code's. But what if the price is one tenth? All is forgiven. It even codes better, so there's nothing to deliberate.

Watching coding agents advance one step at a time stirs something in me. Until around early last year I watched them with a gaze full of gloom and helplessness. What am I going to do for a living? At some point I set the despair down, and the world turned out to be terribly fun. I believe coding, like self-driving, needs AGI to be fully solved, and that LLMs alone can't produce AGI. So for a while, quite a long while, we get to enjoy LLM-assisted coding. The way I once watched Apple, Google, and Samsung launch events ten or so years ago, wondering how upgraded the new phones would be, I now wait for the next AI model, wondering what it will do better, with the same fluttering heart.