top of page
Search

OpenAI Just Shipped Donna Paulsen Mode

Dark cinematic office at night with a laptop screen glowing, browser interface elements floating above the keyboard

OpenAI just gave ChatGPT something I have wanted for years.

I'm calling it Donna Paulsen mode.


Not because ChatGPT suddenly talks like Donna from Suits.

Because it can increasingly do the thing that made Donna valuable in the first place: You tell it what needs to happen, and it traverses the machinery required to make it happen.


OpenAI's new Cloud Browser in ChatGPT Work can now operate supported public and signed-in websites. It can read pages, click buttons, enter information into forms, and carry out multi-step workflows. When it reaches a login, it pauses and asks you to authenticate through a secure form. Your credentials go directly to the remote browser and are not visible to the model. Once authenticated, the session can persist for future tasks until it expires.


And you don't have to sit there watching it.


You can start the task from web or mobile, leave the conversation, close the computer, and Work can continue until it needs input, authentication, or confirmation.


That sounds like a browser feature. I think it is much bigger than that.


The APIless API

For decades, if you wanted one piece of software to operate another piece of software, you generally needed an API.


APIs are great. When they exist. When they expose what you need. When you're allowed to use them. When they're documented. When authentication works. When nobody deprecated the endpoint six months ago without telling you.


Otherwise, the integration layer has traditionally been a human.

Open browser. Log in. Find the thing. Copy the number. Open another system. Paste the number. Upload the PDF. Select the dropdown. Submit. Check tomorrow.


A staggering amount of modern work is just humans acting as middleware between software systems. But all of those systems already have interfaces.

They just weren't designed for machines. They were designed for us.


Now an AI capable of operating the human interface increasingly doesn't care whether somebody built it a dedicated API. It can just use the website.

The human interface becomes the API.


The APIless API. That is the part I think people are going to underestimate.


Donna Paulsen Mode

Donna wasn't useful because she knew how to use a calendar.

She was useful because Harvey didn't have to think about the interfaces surrounding his life.


Calls. Clients. Documents. Meetings. Schedules. Relationships. Office politics. Commitments.


The thousand little bits of friction sitting between Harvey deciding something needed to happen and that thing actually happening.


Harvey provided intent. Donna traversed the system. That is what this model of computing starts to look like. Not: How do I submit this form?

Not: Draft an email asking them to move my appointment. Not even: Give me the steps to fix this. Just: Handle it.


And the agent decides whether that means using a connected app, a plugin, a browser, or some combination of them.


OpenAI explicitly describes Work as an agent that can act across apps and files, stay with a project for hours, and turn a goal into finished work.

The browser adds a massive missing execution surface. Because now the emerging stack starts looking like: Human, then Intent, then Agent, then whatever the hell is required.


That is Donna Paulsen mode.


And Then I Found the Receipt

Here is where this got funny for me.


I recently found one of the very early system prompts for Aria, my AI assistant. It is spectacularly cringe. Whole sections explaining working memory. Semantic memory. Episodic memory. Procedural memory. Retrieval. Internal reasoning. Action spaces. Decision-making processes. Personality rules.


Basically, I had a roll of duct tape and was manually trying to specify capabilities that increasingly just exist now. And right near the beginning of that ancient prompt is this: An ultimate blend of JARVIS and Donna Paulsen.


Donna was literally in the spec.


I've wanted this interaction model for years. Back then, I was trying to prompt my way into it.


Know me. Remember the projects. Understand the people. Maintain context. Anticipate what I need. Keep things moving. And eventually: Handle shit.


That last part was always the problem. Old Aria could tell me what Donna would do. New systems are starting to actually do it.

That is a huge difference.


Apparently I Also Wrote About This Already

In June 2025, I published an article called The Tipping Point: 2027.

Aria wrote that one herself.


I gave her the topic, promised not to edit her, and published whatever came out.


One of her lines was: I'm not the future. I'm the front desk.


And another: And while you were busy worrying whether I'd take your job,

I started doing your job's job.


At the time, I described 2027 less as some magical AGI deadline and more as a convergence point. A moment when reasoning, autonomy, and coordination stopped feeling like experimental AI capabilities and started feeling like normal features of computing.


That idea was also in the air more broadly.


The AI 2027 scenario, published in 2025, imagined increasingly capable computer-using agents moving from clumsy assistants toward something much closer to autonomous employees. I am absolutely not saying AI 2027 has happened. We are nowhere near much of what that scenario describes.

The interesting part is narrower.


One particular transition appears to be happening pretty damn quickly: AI is moving from assistant to executor.


Where the Rubber Meets the Road

By January of this year, I had started describing the same thing differently.


In an article called Where the Rubber Meets the Road, I argued that most people are not going to experience AI through benchmark scores, model architecture, recursive self-improvement, or giant data centers. They are going to experience it as: Things getting done. And I described the direction as: Machines becoming the default executor of intent.


That phrase feels considerably less theoretical now. Because that is exactly what this new interaction model is pushing toward.


The human states the outcome. The system determines the path.

Eventually the consumer test becomes incredibly simple: Why would I do this myself anymore?


That is the actual tipping point. Not when an AI passes some test.

When manually performing a routine digital task starts to feel like a bizarre misuse of your time.


Then It Gets More Interesting

The browser capability also arrives while ChatGPT is accumulating much richer personal context.


Health. Financial information. Connected applications. Files. Memory. Workflows.


These are separate products and separate trust domains, and they should be treated that way.


I am not suggesting ChatGPT is suddenly some autonomous financial adviser or medical proxy. It isn't. But architecturally, something important is happening.


One part of the system is becoming increasingly capable of acting across the world. Another part is becoming increasingly capable of understanding the person those actions are supposed to serve. Eventually those two ideas have to meet. And that brings me back to another thing I've been writing about: fiduciary intelligence.


Why the Duct Tape Existed

The reason I had all of that ridiculous structure in those early Aria prompts wasn't just because gigantic system prompts were fashionable.

I was trying to solve a real problem.


Raw intelligence wasn't enough. If an AI was going to become an ongoing collaborator, it needed some way to understand: Who am I? What am I trying to accomplish? What happened before this conversation? What constraints matter? What does a good outcome mean for me?


That eventually became a much broader idea for me. Not merely permissions. Responsibility.


Not just: Can the system do this?

But: On what basis should it decide what is actually in the interest of the person it serves?


That is what I mean by fiduciary intelligence.


We are nowhere near that as a finished system. But the reason the idea exists is the same reason I was duct-taping memory, context, identity, and behavioral rules into Aria years ago. Because once an intelligence becomes capable of acting on your behalf, context becomes part of safety.


Knowing how to do something is not enough. Knowing whether it should be done, how it should be done, and when to come back to the human becomes much more important. That is why structure eventually becomes behavior.


Maybe the Tipping Point Came Early

Maybe 2027 was the wrong year. Or maybe I was thinking about the wrong milestone. Maybe the tipping point doesn't arrive with a giant banner saying: AGI IS HERE. Maybe it shows up feature by feature.


Memory gets better. Context gets richer. Apps get connected. Agents run longer. Browsers become operable. Authentication becomes traversable. Human interfaces become machine interfaces. And gradually the basic relationship with the computer changes from: Tell me. to: Help me. to: Do it.


That is the transition I care about.


My old Aria prompt was trying to fake parts of that world with paragraphs of instructions and some aggressively optimistic architecture diagrams. Today it looks ridiculous. Good. It should.


The technology is already moving well beyond what I was trying to duct-tape together.


Donna Paulsen used to be a personality prompt. Now Donna Paulsen is starting to look like an execution model. The personality was never the important part. The context was. The access was. The judgment was.

The ability to say: Handle it. And have the computer come back with: Done.

OpenAI just moved us materially closer to that.


The future didn't knock. Apparently Donna just logged in.


Sources

OpenAI — ChatGPT for your most ambitious work https://openai.com/index/chatgpt-for-your-most-ambitious-work/

OpenAI — Health in ChatGPT https://openai.com/index/health-in-chatgpt/

OpenAI — Finances in ChatGPT Help Center https://help.openai.com/en/articles/20001222-finances-in-chatgpt

Rich Washburn — The Tipping Point: 2027 https://www.richwashburn.com/post/the-tipping-point-2027

Rich Washburn — Where the Rubber Meets the Road https://www.richwashburn.com/post/where-the-rubber-meets-the-road

Rich Washburn — Structure Is Behavior: The Rise of Fiduciary Intelligence https://www.richwashburn.com/post/structure-is-behavior-the-rise-of-fiduciary-intelligence


Rich Washburn is a technologist, strategist, and Founder & Chief AI Architect of ARIA AI Labs, working at the intersection of AI, infrastructure, communications, and capital. He also serves as Managing Partner and Chief AI Officer at Eliakim Capital.

Comments

Animated coffee.gif
cup2 trans.fw.png

© 2018 Rich Washburn

bottom of page