It’s 12:40 on a Tuesday. You have eleven minutes before your next meeting, your laptop is open to four tabs you can’t close, and you’re hungry in that specific way where deciding what to eat feels harder than the work itself. So you pick up your phone, open Messages like you’re texting a friend, and type: “order my usual protein bowl to the office.”
Then you go back to your meeting. The food shows up.
That’s the experience DoorDash is piloting with a new AI agent that takes food orders over text inside Apple Messages. A US waitlist is open now. And if you’ve been wondering what people mean when they say “AI agent,” this is one of the cleanest examples you’ll find.
What makes this an agent and not just a chatbot
The word “agent” gets thrown around loosely, so let’s be precise about it. A chatbot answers you. An agent does something for you.
When you text that request, the system doesn’t just reply “Great choice!” and hand you back a link. According to DoorDash, it searches local restaurants, scans menus, and builds a suggested cart based on what you asked for. Three separate jobs, chained together, triggered by one sentence of ordinary human language.
Think about what that sentence actually contains. “My usual” implies memory of past behavior. “Protein bowl” is a category, not a menu item, so something has to translate it into real dishes from real restaurants near you. “To the office” is an address you never typed. A person could sort all that out in a few seconds without thinking. Software historically could not, which is why we’ve spent fifteen years tapping through apps instead.
The interface is the interesting part
The technical work here is impressive, but the choice I keep turning over is where DoorDash put it: inside Messages. Not a new app. Not a voice assistant. Not a chat window buried behind two taps in an existing app.
Messages is probably the app you use most without ever thinking about it. There’s no learning curve, no onboarding flow, no feature tour. You already know how to type a sentence to someone and expect a result.
That’s a quietly big deal for anyone who finds new technology exhausting. The hardest part of adopting a new tool usually isn’t the tool, it’s the mental overhead of learning where the buttons live. Putting an agent in a text thread removes nearly all of that. The skill you need is one you learned years ago.
Where ordering stops being shopping
Food delivery apps are built to be browsed. Photos, categories, promoted listings, a carousel of things you weren’t looking for. That design isn’t an accident, it’s how these platforms work. Browsing produces bigger carts.
Texting an order skips the browsing entirely. You state intent, you get a cart. It’s less like shopping and more like delegating, which is a different relationship with an app than most of us are used to having.
What to keep an eye on
I’m genuinely interested in this one, and I also think the honest take includes some open questions. Based on what’s been announced, here’s what I’d want to understand before I trusted it with my lunch money.
- How much does it decide for you? The agent builds a suggested cart. The gap between suggesting and ordering is where trust lives. A confirmation step before anything is charged matters a lot.
- What does “my usual” mean to the system? Memory is what makes this feel magical, and memory means stored history about what you eat, when, and where you are. Worth knowing how that’s handled.
- What happens when it misreads you? Natural language is messy. “Something light” means different things to different people. Fixing a wrong order needs to be as easy as placing the right one.
- Who picks the restaurant? When you browse, you choose. When an agent chooses, the logic behind that choice becomes a question you can’t see the answer to.
None of these are reasons to be suspicious. They’re just the questions that come with handing off a decision, and they apply to every agent you’ll meet, not only this one.
Why a lunch order is a reasonable place to start
If you want to get people comfortable with software acting on their behalf, food is a smart first step. The stakes are low, the feedback is immediate, and the task repeats often enough that small conveniences compound. If the agent gets your bowl wrong, you’re mildly annoyed and out some money. You learn where the limits are without much risk.
That matters because the same pattern is heading toward things with real consequences. Booking travel. Managing calendars. Moving money. The habit of describing what you want in plain language and letting software handle the steps is the habit being built here, one lunch at a time.
For now, it’s a waitlist and a text thread. But if you’ve been curious what an AI agent actually does, a protein bowl arriving at your desk is a surprisingly good answer.
🕒 Published: