We’ve all done it. You ask an AI to plan a trip. It gives you a flawless itinerary — flights, hotels, restaurants, the works. Then you spend the next forty minutes manually typing all of that into six different apps.
The problem with AI isn’t that it can’t think. It’s that it can’t do anything.
Qi Junyuan knows this better than most. He was the person behind ByteDance’s Doubao PC client — the desktop version of one of China’s most-used AI assistants. He watched millions of users ask questions, get answers, and then go do the actual work themselves.
Now he’s left ByteDance. He’s raised two rounds of funding from IDG and other top-tier dollar funds. His company is valued in the hundreds of millions. And he’s building something called Today AI — an operating system with, as he puts it, “a mind of its own, focused on you.”
Here’s the twist: “Today” isn’t a new name for him. A decade ago, at his previous company Teambition, he launched a productivity app called “Today.” It was a to-do list tool. It failed and shut down.
Same name. Radically different ambition. This time, Today isn’t about reminding you what to do — it’s about doing it for you.
The vision is what the industry calls a GUI Agent: an AI that can see your phone screen, understand what’s on it, and take action. Not “here’s how to book a flight” — actually booking the flight. Filling the forms. Selecting the seats. Confirming the payment. Autonomously, across multiple apps, in sequence.
This is where most coverage stops. “AI that operates your phone!” Cool story. But here’s what nobody’s talking about:
The real bottleneck for AI agents isn’t intelligence. It’s permission.
Think about it. Every app on your phone — your banking app, your ride-share app, your food delivery app — is a walled garden. They’re designed for human fingers, not AI intermediaries. They have CAPTCHAs, session timeouts, anti-bot protections. They don’t want an AI logging in on your behalf.
For Today AI to work, one of two things has to happen. Either app developers voluntarily open up their APIs and permission layers to let AI agents operate within their ecosystems — or the AI has to brute-force its way through interfaces designed to keep it out.
The first option requires cooperation from an industry that has zero incentive to cooperate. Why would Uber let an AI OS book rides when Uber has its own in-app experience to protect? Why would a bank let an AI move money when the liability is astronomical?
The second option — brute-forcing through GUI manipulation — is technically possible today. But it’s fragile. One app update, one redesigned button, one new pop-up, and the agent breaks. You can’t build a reliable operating system on top of interfaces you don’t control.
Every app developer faces a choice: open up and become a backend service for someone else’s AI, or get bypassed entirely.
This is the real battle. Not model versus model. Not GPT versus Claude versus Doubao. The fight that matters is between AI operating systems and the app ecosystems they’re trying to disrupt.
Qi Junyuan understands this — or at least he’s betting that he can find a way through. Today AI is already in testing across Mac, iOS, and Android. It supports email, Google, and Apple login — a clear signal that this isn’t a China-only play. The TestFlight build was updated on July 20th. The team has been building for at least three months, probably longer.
But the valuation — hundreds of millions of dollars for a product still in closed beta — tells you what investors are really betting on. They’re not funding a chatbot. They’re funding the possibility that someone cracks the permission problem.
The company that solves the integration layer wins the AI era. Not the company with the smartest model.
We’ve seen this movie before. The last platform war was won not by the best hardware or the best software, but by the company that controlled the distribution layer. Apple didn’t win because the iPhone was technically superior — it won because it controlled the App Store, the toll booth of the mobile era.
Today AI is trying to build the next toll booth. Not an app store, but an agent layer — a system where AI operates everything, and every app either plays along or gets rendered invisible.
Will it work? The honest answer is: nobody knows. GUI agents are still in their infancy. The reliability isn’t there yet. The permission infrastructure doesn’t exist. The app ecosystem hasn’t decided whether to fight or adapt.
But here’s what I do know: the next trillion-dollar AI company won’t be the one that writes the best prose or generates the prettiest images. It’ll be the one that bridges the gap between thinking and doing.
Chatbots answered our questions. The next wave of AI will answer our to-do lists.
And Qi Junyuan — the man who once built a to-do app that failed, and is now building an AI that could make to-do lists obsolete — is betting everything that the future doesn’t talk. It acts.
FAQ
Q: If app developers won't cooperate, can't the AI just operate apps visually without permission?
A: Technically yes — GUI agents can click buttons and fill forms like a human would. But it's fragile. One app update, one new pop-up, one redesigned interface, and the agent breaks. You can't build a reliable OS on interfaces you don't control. It's a hack, not a foundation.
Q: What does this mean for regular users?
A: If Today AI or similar products succeed, your phone assistant could actually book flights, fill forms, and manage tasks end-to-end without you touching each app. If they fail, we're stuck with chatbots that give great answers but can't lift a finger to help.
Q: Is a hundreds-of-millions valuation for a closed-beta product insane?
A: On paper, absolutely. But investors aren't valuing the product — they're valuing the option on who wins the integration layer. If someone cracks the permission problem, they become the toll booth for all AI-agent interactions. That's a platform play worth betting on, even if the odds are long.