OpenAI DevDay: 'o' Agent | Pixel Audit

MayaOkay, so I'm sitting here with my little grading sheet again, and I genuinely don't know what column half the stuff goes in.

GrantThe spread you taught already? We haven't even seen the keynote.

MayaGrant, it's DevDay. You know I bring the scorecard.

GrantWe've used it on every Apple event this year, every Google event.

MayaRight. Pass, fail, pass, fail, clean little check boxes.

GrantAnd today?

MayaToday, the whole category barely has rules yet. Agentic AI, this always-on assistant stuff, it-- there's no design convention to hold it against.

GrantSo we're grading blind.

MayaBasically.

GrantI mean, I love a rubric with no rubric.

MayaIt's like judging a gymnastics routine nobody's ever performed before.

GrantSure, but somebody still has to hold up a number.

MayaExactly. So that's us today.

GrantI'm Grant.

MayaI'm Maya, and this is the show where two designers argue about whether a product actually works.

GrantTwo hooks before we even start, what's actually shipping versus what's just a rumor with a good headline.

MayaAnd whether anyone at OpenAI has thought about trust as a design problem at all.

GrantOr if it's just vibes and a progress bar.

MayaWe'll find out.

GrantOkay, so the actual event, OpenAI's own announcement page just says, "DevDay 2026, September 29th, San Francisco." That's it. That's the whole confirmed part. That's so funny.

MayaThey don't even tease what's on stage? Nope, just the date and a city.

GrantWell, CNBC's live page is doing more work than OpenAI is, honestly.

MayaThey're framing this whole keynote against the safety heat OpenAI's been getting. Right.

GrantAnd Altman posted on X that they'd found a new thing. And Altman posted on X that they'd found a new thing.

MayaA new thing, very specific.

GrantExtremely specific. Keynote's 10:00 AM Pacific, so it's already going by the time we're recording this.

MayaWhich is wild. We're basically live grading something we can't see yet.

GrantHonestly, I kinda like grading without spoilers for once. Usually by showtime, some leaker's already told us the whole runway.

MayaToday, we've just got a date, a city, and a guy on X saying he found a new thing. That's the whole briefing packet.

GrantThat's the job today.

MayaSo before the actual keynote drops, let's talk about what people were expecting walking in. There's a piece from this morning saying DevDay could bring more than 20 product announcements.

Grant20? That's a lot to cram into one keynote.

MayaYeah, and it's not just software, new AI hardware, ChatGPT upgrades, and get this, a rumored subscription tier at $500.

GrantA month?

MayaThe piece doesn't even nail that down. It's just floating out there as a rumor.

GrantOkay, but I wanna flag, that whole article is pre-keynote. None of that is confirmed yet.

MayaRight. It's expectations, not receipts.

GrantAnd honestly, expectations not receipts is basically our disclaimer for the whole episode today.

MayaWe should put that on a T-shirt.

GrantWhich actually tells you something about how we should grade this. We need two columns today. Column one, does this new agent category even have UX conventions to measure against?

MayaAnd column two?

GrantColumn two, is OpenAI making those conventions up live on stage in real time?

MayaBecause if they are, we're not grading a product, we're grading a first draft.

GrantExactly the problem.

MayaCan I push on the safety framing for a second? Because CNBC didn't just mention it as color, they built the whole live page around it.

GrantRight, which is a weird position for a UX show to grade from. We don't normally ask, "Is the company under pressure?" We ask, "Does the button work?"

MayaBut if the pressure's the backdrop, doesn't that change what counts as a pass? Like, showing restraint might actually be the right call today, even if it makes for a boring keynote.

GrantSure, except restraint and we don't have a UX answer yet look exactly identical from the outside. That's exactly the trap.

MayaFair. And that $500 tier rumor, even if it's just floating out there, it tells you something about what they think the ceiling is.

GrantIt tells you what a reporter thinks the ceiling is. Big difference.

MayaOkay, sure, noted. But if any of that Twenty plus number is real, that's a lot of surface area for us to grade today.

GrantWhich is Honestly my worry walking in. Twenty announcements in one Keynote Usually means Five get real design attention, and the other Fifteen get a slide and a shrug.

MayaSo part of our job today is figuring out which pile each thing lands in.

GrantExactly.

MayaOkay. So with all of that as the backdrop-

GrantThere's one Rumored in that pile that's bigger than the other Nineteen combined

MayaThe Agent.

GrantThe Agent. Everybody's been chasing this one thing specifically, and it's still not Confirmed. Okay, so digging into this, the leaks say OpenAI's preparing a new assistant called O for release, and it's supposed to be Always-On.

MayaAlways-On, meaning it doesn't shut off when you close the tab.

GrantRight. It just keeps going.

MayaWhich is exactly the kind of thing that needs a design plan before it ships, not after.

GrantSo there's a whole trending thread on X people keep passing around, an orb teaser, and somebody spotted a leaked line on the Pro upgrade screen that just reads, "O, your always on assistant."

MayaWait, an orb? Like an actual glowing ball?

GrantThat's the read going around, yeah. People stitching screenshots together. Of course they are.

MayaBut that's all community noise.

GrantRight. And then there's a separate tracker I came across, a report that's been updating in real time, careful about what's confirmed versus not. And as of its last check before the keynote even started, they hadn't found a single official OpenAI announcement of O. Nothing on the record.

MayaSo it's rumor stacked on rumor.

GrantWhich is the part I want to sit on. Say it's real. Say it's always on. What does that look like on your phone? Does it notify you? Does it just listen? Where's the off switch?

MayaI mean, Siri and Alexa figured some of that out years ago. Wake words, mute buttons.

GrantSure, but those are single purpose. This thing's supposed to act, book stuff, send stuff, touch your calendar. Nobody shipped a consent screen for an agent that does things on your behalf and just keeps running in the background.

MayaSo you want a consent screen before it even ships?

GrantI want to know there is one somewhere. Right now, we've got an orb and one line of leaped copy. That's marketing, not a settings page.

MayaOkay, walk me through what you'd actually want to see then. Not the vibe, the actual screen.

GrantSomething like what location sharing apps already figured out. A persistent bar at the top that just sits there the whole time saying, "This app knows where you are right now." Not a one-time pop-up you tap through once and never see again.

MayaSo for O, that'd be something always visible saying exactly what it's currently allowed to touch.

GrantRight. Calendar, email, whatever task is midway through, and a real off switch sitting right next to it, not buried three menus deep in settings where nobody finds it.

MayaWhat about the notification side? Does it ping you every single time it does something, or does it just log it quietly somewhere?

GrantThat's the other half of the problem nobody solved. Too many pings, everybody mutes the whole thing in a week. Too quiet, and it's just rearranging your life while you're not looking.

MayaWhich is kind of why Siri and Alexa never had to fully solve this. They sit there waiting for a wake word. They don't go do things on their own while you're asleep.

GrantExactly. They never needed a constant status indicator because they weren't constantly doing anything. This thing is supposed to be different by design, and that's the exact piece I haven't seen a single screenshot of. Not the orb, not the leaked line, the actual status view.

MayaFair. Although if it's this early, maybe nobody's finalized the design yet.

GrantMaybe, but that's exactly the gap I keep circling back to.

MayaYeah. Yeah.

GrantAnyway, the agent's the headline nobody's confirmed. There's something underneath it though that's a little more concrete.

MayaThe desktop app thing?

GrantYeah. Reportedly, OpenAI isn't just bolting O onto ChatGPT. They're reportedly folding it into some kind of bigger unified app. Exact scope's not pinned down.

MayaWait, merging things? When?

GrantThat's the part nobody's pinned down. Could be part of today, could be totally separate.

MayaSo now we've got an Unconfirmed agent sitting on top of an Unconfirmed merger.

GrantWhich is exactly where I wanna go next because that's where the actual handoff problem lives. Okay, so buried under all the agent hype, there's this Wikipedia entry on Codex that's actually pretty specific. It says OpenAI is combining the desktop app, Codex, and Atlas into one super app.

MayaWait, Atlas too? I thought Atlas was the browser thing.

GrantIt is. So you'd have Chat, the coding agent, and a whole browser all living under one roof.

MayaOkay, but that actually makes sense to me. Anthropic's out here shipping Claude everywhere, in the terminal, in the browser, in your IDE. If you're OpenAI, you can't have three separate apps competing with that.

GrantSure. From a spreadsheet, it makes sense. One SKU, one login, one thing to market.

MayaRight?

GrantBut from a screen, that's three totally different design languages getting shoved into one window.

MayaYou say that like it's automatically bad.

GrantI mean, is it not? A chat interface and a browser and a code editor are not the same shape of tool. You don't use them the same way. You don't think in them the same way.

MayaSo what? You'd rather they stayed three separate downloads people forget which one to open?

GrantNo, I'd rather they actually designed one flow instead of duct taping three roofs onto one house.

MayaOkay. Harsh.

GrantHere's what the brief actually promises, and I wanna hold them to it. Start a task in Chat, hand it off to Codex to keep working, then check the result in a dashboard.

MayaThat's the whole pitch, yeah. Chat to Codex to dashboard.

GrantOn paper, that's kind of elegant. In practice, that's three separate surfaces the user has to track a single task across.

MayaBut isn't that just how work already happens? You draft something in one app, hand it to another, check it in a third.

GrantSure, except normally you're the one doing the handing off. You know where the thing went.

MayaAnd here, the agent's doing it for you.

GrantRight. So if it breaks or stalls or does the wrong thing halfway through Codex, where do you even go to find out? Which tab?

MayaOkay, that's fair. I don't actually know the answer to that.

GrantUh, neither do I, and that's the test I wanna run on this whole merger. Does folding three apps into one window solve that, or does it just move the clutter up one level?

MayaI think it genuinely could go either way, depending on whether the dashboard is actually a single source of truth or just another tab you have to remember exists.

GrantExactly the thing. If it's one shared state across Chat, Codex, and the dashboard, great. That's a real merger. If it's three apps with a shared frame around them-

MayaIt's just a browser with more windows.

GrantRight, and we don't know which one it is yet.

MayaWe do know somebody at OpenAI has already been pretty blunt about how the desktop app looks right now, though.

GrantOh, this is the part I've been waiting to get to.

MayaYeah, because if the handoff promise depends on the app actually being clean and legible and the guy running it already called it messy-

GrantThen the super app isn't inheriting a blank canvas. It's inheriting the mess.

MayaOkay, so the redesign confession. This is the one where Brockman just said it out loud.

GrantYeah. AndroidHeadlines ran this back in July. OpenAI's own president called the unified ChatGPT desktop app a kind of mess. His words.

MayaThat's not a leak. That's a guy just being honest on the record.

GrantRight. And he teased a redesign, tab-free, supposedly seamless, coming after the backlash.

MayaTwo months before today's keynote.

GrantWhich is exactly why I don't buy the seamless part yet. You don't fix three stitched together products by promising fewer tabs.

MayaWait, why not? Fewer tabs is fewer tabs.

GrantBecause the tabs aren't the disease, they're the symptom. You stack chat and agent runtime and a coding tool into one shell, and clutter is what falls out. Hiding the tabs doesn't remove the seams underneath.

MayaSure, but give them a second. You don't ship three products merged on day one with a perfect shell. Somebody has to live in the messy version, so the real one gets built.

GrantThat's the interim state argument. I get it. I just don't think kind of a mess from the guy running the company is an interim state. That's a confession.

MayaFair. Okay, before I forget, we should do a quick round. Red flag or green flag?

GrantGo.

MayaTab-based navigation for an agentic product.

GrantRed. Tabs assume you're switching between static documents. An agent that's doing things behind the scenes needs status, not tabs.

MayaAgreed. Red for me too. Next, notification design for an Always-On agent.

GrantAlso red, and it's not close. Nobody's shown us what that notification even says yet.

MayaYellow for me. It's red if they get it wrong, but at least it's a real design problem somebody has to solve, not an afterthought.

GrantFine. Yellow leaning red. Last one, dashboard as an inspection panel for what the agent already did.

MayaGreen. That's the one piece of this whole thing that has actual precedent. Build tools have had inspection panels forever.

GrantGreen for me, honestly. First green flag of the segment.

MayaLook at that. We agree on something.

GrantDon't get used to it.

MayaSo zoom out for a second. We keep grading tabs, notifications, dashboards, but none of that's really the question, is it?

GrantNo. The question is whether any of this is even the same category of thing we've graded before.

MayaRight. Like a redesign of a chat app is one kind of problem. You move buttons, you fix contrast, you ship it.

GrantBut an agent that acts on its own across three surfaces without you watching every step, that's not a layout problem. That's a whole different kind of interface.

MayaAnd if it's a different kind of interface, our usual scorecard might not even apply.

GrantWhich is the actual thing we should be asking about DevDay. Not is the app clean? But is this even a UI in the sense we've meant that word for the last sixty years? And so NNg, the usability people, actually put a number on this, sixty years.

MayaSixty years since what?

GrantSince the last new way of interacting with a computer got invented. Jakob Nielsen's own piece calls AI the third paradigm, command line, then graphical windows and icons. Now this.

MayaWhat's this exactly?

GrantHe calls it intent-based outcome specification. You tell it what you want, not how to do it.

MayaThat's a mouthful for "I type to sentence."

GrantSure, but think about what it replaces. Clicking through menus, learning where buttons live. That's sixty years of convention we all just absorbed without noticing.

MayaRight. And today's keynote is supposed to be built on a paradigm that old and that new at the same time.

GrantWhich is exactly the tension, yeah.

MayaBecause windows and icons came with rules. You knew a button was a button because it looked pressed. Nobody's agreed what an agent quietly acting on your behalf is supposed to look like.

GrantThere's no equivalent of the pressed button shadow for "I did a thing while you weren't looking."

MayaExactly. And that's the actual question under all of it. Is trust a design problem anyone over there is solving? Or is it just getting waved through with words like smooth and effortless?

GrantWords that mean nothing until you can point at the screen and say, "Here's where it tells you what it did."

MayaRight. Show me the moment, not the tagline.

GrantAnd today we didn't really get that moment. We got rumors, a leaked screen, a merger nobody's confirmed the timing on.

MayaWhich sixty years in is kind of the whole story. The interaction model changed. The proof you can trust it didn't ship yet.

GrantSo where does that leave the scorecard? Honestly, incomplete. Confirmed, the keynote happened, the merger's reported. Unconfirmed, almost everything about how it actually behaves day to day.

MayaThe agent piece stays a rumor until somebody at OpenAI shows the off state, the notification, the log of what it touched.

GrantAnd the redesign is an admitted mess with no finished version to grade.

MayaSo it's not a pass or fail episode, it's an incomplete.

GrantWhich is fine. That's an honest grade. But listeners can still use this today.

MayaYeah. Here's the takeaway. Anytime a product pitches a seamless handoff, whatever the product, ask one question.

GrantWhere does it tell the user what the system just did on its own?

MayaIf there's no answer to that, it's not seamless, it's just quiet.

GrantQuiet's not a compliment when it's your calendar getting rearranged.

MayaNo, it is not.

GrantSo that's the scorecard for today. Two confirmed facts, three open questions, and one sixty-year-old paradigm that still hasn't earned the trust part.

MayaWhich is a pretty good place to leave it.

GrantOkay, one steal this each, then we're out. Mine, before you trust any agent flow, check if it labels which mode you're actually in right now. Chat, autonomous, view. If you can't tell at a glance, that's a design failure, not a user error.

MayaOoh, good one. Mine's simpler. Is there a visible log you can scroll to to see what the thing did without you? Not a settings menu buried three clicks deep. Front and center.

GrantRight. Because if you have to go dig for it, it's not transparency, it's plausible deniability.

MayaOkay, official business time.

GrantIf you spot a UI that bugs you this week, screenshot it and tag us. We might grade it live.

MayaSubscribe so you catch next week's audit, whatever ships out of this keynote.

GrantAnd if this changed how you look at your own product screens, leave us a review. Genuinely helps.

MayaTotally. All right, Grant, go check if your dashboard actually tells you what it's doing.

GrantWorking on it. Talk soon, everybody.