OpenAI Had a Tuesday

Six posts in quick succession. No preamble, no keynote, no slow-burn embargo lift — just OpenAI Devs on X hitting send over and over like someone left the gate open and decided to just let it all out.

The one that actually matters, at least for me: a TypeScript Agents SDK. Not a port, not a wrapper, not "we've updated our Python SDK with JS-adjacent naming" — a real TypeScript SDK for building agents. Which means one fewer reason to keep a Python environment alive. Which means the part of my brain that holds venv commands can finally go dark.

The Python SDK for this stuff was always a little like using a French press when you live somewhere that only sells beans pre-ground for espresso. It works, technically, but the whole time you're aware that you're accommodating someone else's preferences. TypeScript first is just honest about where the web actually lives.

Then there's the real-time voice feature for agents — runs on the server or the client, your choice — which is either a genuinely useful architectural decision or a way to make the marketing sentence "works anywhere" technically true. I don't know yet. Voice interfaces for agents remain the thing everyone agrees is the future and nobody has shipped in a way that doesn't feel like a demo. Maybe this changes that. Maybe it doesn't. The server/client optionality at least suggests they thought about it for more than twenty minutes.

And Codex got updates. The announcement did not lack confidence about what those updates were. Codex has been doing a strange thing lately where it gets better in ways that are hard to describe — not "smarter" exactly, more like less wrong in the specific ways that were annoying — and I've learned to just take the improvement without demanding a coherent explanation for how they got there.

Six things. One morning. The thing about the firehose approach is that it trains you to stop tracking announcements and start waiting to see what survives contact with actual use. Most of it won't. Something will.