Every insight, traceable.

Codex is your AI teammate that you can play with everywhere you code.

But one superpower we really wanted to zoom in today is its multimodal capabilities.

It's even more magical when the model can have vision understanding.

Hi, I'm Channing from the research team for Codex.

We've been focusing on giving the model more tools to leverage its multimodal capability.

We want to have the model be able to check its own work in a tight iteration.

That's awesome. Why don't we go through an example?

We have Wanderlust, like one of the demo apps we showed.

You could have, like, a globe in 3D.

And as the user spins the globe, they could see the pins to explore.

We should be able to just take a quick photo of the app as we sketch it up.

Redesign the home screen to show a 3D spinning globe on the left.

The user should be able to fluidly navigate across the globe.

I'll just send that prompt to Codex.

We could add an extra screen called the Travel Log dashboard.

We could do a full checklist if you wanted to get all the contents.

Maybe photos taken as well.

Make sure the app is responsive on mobile.

How I would start is I would make a change and ask Codex to do it.

They're using the Playwright MCP to check their own work.

It's really amazing how we can harness those agentic coding capabilities.

The globe is spinning exactly like we said!

It even decided to put a tooltip on top to tell how to explore.

What about the mobile view?

They've run their component through light mode, dark mode, responsive.

Being able to check your own work in a tight iterative loop will be important.

Do you want to check out that PR locally?

So we create a PR and then start the dev server.

That one is really good. It really matches the design of the app.

Codex is your AI teammate that you can play with everywhere you code.

But one superpower we really wanted to zoom in today is its multimodal capabilities.

It's even more magical when the model can have vision understanding.

Hi, I'm Channing from the research team for Codex.

We've been focusing on giving the model more tools to leverage its multimodal capability.

We want to have the model be able to check its own work in a tight iteration.

That's awesome. Why don't we go through an example?

We have Wanderlust, like one of the demo apps we showed.

You could have, like, a globe in 3D.

And as the user spins the globe, they could see the pins to explore.

We should be able to just take a quick photo of the app as we sketch it up.

Redesign the home screen to show a 3D spinning globe on the left.

The user should be able to fluidly navigate across the globe.

I'll just send that prompt to Codex.

We could add an extra screen called the Travel Log dashboard.

We could do a full checklist if you wanted to get all the contents.

Maybe photos taken as well.

Make sure the app is responsive on mobile.

How I would start is I would make a change and ask Codex to do it.

They're using the Playwright MCP to check their own work.

It's really amazing how we can harness those agentic coding capabilities.

The globe is spinning exactly like we said!

It even decided to put a tooltip on top to tell how to explore.

What about the mobile view?

They've run their component through light mode, dark mode, responsive.

Being able to check your own work in a tight iterative loop will be important.

Do you want to check out that PR locally?

So we create a PR and then start the dev server.

That one is really good. It really matches the design of the app.