Agents for ALL

OpenAI released ChatGPT Work yesterday and everyone has the same response:

What the hell is this?

Fair!

My Whatsapp chat is full of peeved people. And…I get it.

OpenAI now has Chat, Work and Codex inside one app. Then you can choose a model. Then you can choose an effort level. Then some options appear on desktop but not mobile, or on mobile but not web, depending on whether the rollout gods like you today.

It's a bit of a mess!

BUT…underneath the rubbish naming, something very important just happened. And I think most people missed it.

For 1 BILLION people just ChatGPT stopped being somewhere you ask questions.

One App, Three Doors

OK so what do we actually have now.

The simple version:

• Chat: ask a question and get an answer.

• Work: give it an outcome and let it plan, act and deliver.

• Codex: build software with the full developer controls visible.

Same house. Three doors.

The weird bit is that the walls are a bit flimsy. HOW and WHERE you access these three modes depends on what tool you are using:

• Phone app

• Web interface

• Desktop app

On the Desktop app you switch between Work and Codex. They seem pretty damn similar though - just some visibility differences. You can also pop up a small Chat window for quick questions. But otherwise Chat is gone.

On the web version you can switch between Chat and Work.

On the phone version you have Chat and Work and Remote (to Codex).

Oof. Complex. And annoying to relearn.

That is why power users are moaning. It feels redundant.

But there are millions (nay, 1 billion) people who will happily use ChatGPT every day and will never install Codex, Claude Code or OpenClaw. They hear "coding agent" and do one.

Give them a toggle marked Work though…nice! They’ll get that. And because of that they’ve just been handed agentic AI.

The Unit Of Work Changed

My livestream presentation was made with ChatGPT Work.

I gave it screenshots from Twitter and a long voice note with me blabbering about the launch. It worked for about 12 minutes, spun up three sub-agents, checked official sources, found outside research, outlined the story, made the slides with ImageGen and packaged the lot into a PDF.

I steered it. It did the faff whilst I had my coffee.

If you've already been turning Codex into a personal assistant, that sounds normal. So what? But for a standard ChatGPT user it is a proper change.

Chat has trained us to ask, "How do I make this report?"

Work lets you say, "Make the report. Here are the meeting notes, the source files and the format I need. Bring me back a finished draft."

And that power is now in the hands of everyone.

The Fourth Leap Goes Mainstream

Ethan Mollick talks about four big jumps in modern AI.

GPT-3.5 made the whole thing visible. GPT-4 made it useful for real work. Reasoning models made longer, harder problems possible. Then late last year we got agent systems that could actually go away and complete jobs.

OpenClaw, Claude Cowork and Codex already did this. ChatGPT Work matters because OpenAI has put the same basic behaviour behind a normal-looking toggle in the product everybody already knows.

Distribution matters.

Give It One Real Job

OK how to actually get started with this? This is not about the tech. It’s about your skill level. You need to up your delegation skill.

First up, don't start with a grand plan for an autonomous AI company. Yeesh.

Pick one job you already do that takes about an hour. A report. Meeting notes that need turning into actions. A pile of files that needs analysing. A voice note and screenshots that need becoming a presentation.

Then give Work four things:

1. The outcome you want.

2. The sources it should use.

3. The boundaries it must stay inside.

4. A clear definition of what done looks like.

Your prompting skills still matter because prompting is communication. The better you can explain the result, context and boundaries, the less rubbish you get back. The Prompting Fundamentals playbook still applies. Work just gives those instructions to a system that can keep going for 12 minutes instead of firing back one answer.

Update the app. Find Work. Give it one real job.

To the Task,

Kyle