What Grok Bot Learned From Western Waiters

Why the small things in agent user experience add up, and who's getting them right for now

One of the weird things I’m noticing in agents is something subtle but really smart and helpful. I call it Western waiter training.

Here in Asia you see it a lot at Western-owned restaurants, the ones that have taken the time to train their staff to reduce errors and improve customer service. Say you order a pizza and ask for double mushrooms and no onions, plus one large bottle of water and an order of six garlic rolls. A Western-trained waiter will say, “Got it, so let me repeat your order. Your pizza is going to have double mushrooms and no onions, you ordered one large bottle of water, and an order of six garlic rolls. Did I get that right?”

You confirm it. Then she says, “Great, I’ll put your order in and be right back with your water.”

That’s nice. You’ve confirmed the order is going to be correct, so you can sit back and relax while your food gets ready, because you know there won’t be any mistakes.

Grok Bot got the training

What I recently found is that Grok Bot has been trained in some way, shape or form to do the same thing, and it’s very helpful. It gets your order right, and it also greets you with what feels like a real desire to please you and nail the order exactly how you want it done.

That’s completely missing from OpenAI’s version of an agent, and it’s definitely missing from Claude’s.

The same pizza, three ways

Let me show you what Claude does.

Claude says, “Okay, I’m working on your pizza. By the way, you’re going to need to go get the mushrooms out of the refrigerator. As for the onions, I’ll try my best, but I need you to connect to our kitchen directly and tell them about the onions. So pull out your phone, call this number, and make the kitchen aware of your onions. Oh, and I’m not sure I want to put that much garlic on your garlic rolls, because it might not be safe, and I don’t know if I have permission to do that.”

OpenAI basically says, “I’m going to work on your pizza,” and then it comes back. Most of the time the order is pretty much right. But it doesn’t give you any confirmation, and it doesn’t make you feel like it’s eager to help.

That’s the main difference between the three right now. I think Grok Bot understands a lot of little things at once, and they compound into a better user experience. It’s the same as sitting down at a restaurant and getting that kind of service.

The little things in AI agent user experience add up

Grok Bot also does some really small things that make a big difference.

If you have an API key, you don’t have to figure out where to put a .env file somewhere on your machine. It serves up a nice little card. You copy and paste the key, click save, and it saves it securely for you. It’s so nice. Claude definitely doesn’t do that, and Codex doesn’t either.

Another one: like a good Western waiter, Grok Bot tells you right off the bat when you’re going to need to log in to something. And instead of only telling you what to do, it gives you a link. You click it and land exactly on the login page or approval page you need to move forward in the workflow.

What does Claude do? It gives you a wall of text, and you have to work out where to log in and what domain to use. It tells you what to do without pointing you anywhere or giving you a link that would make your life so much easier.

These little things are really rough and crude in Claude right now. Even though it has a better model and a great harness, the experience is just crude. That’s especially true in Projects, which I get is pretty new, something they just recently launched. But Projects is really rough when it comes to navigating all your threads and feeling like you’re actually in control of them.

Then there are the permissions. You have to type in these stupid phrases like go push server 246. What? I should just be able to click Approve or Go. Instead it wants this ridiculous permission theater where you type everything out, and it’s a lousy experience.

My ranking right now

So I want to give props to Grok Bot. They’ve figured out the Western waiter approach, and they’ve polished up a lot of the rough edges in the user experience. Claude is simply dumping those rough edges on the table and expecting you to finish the job.

OpenAI comes in second out of the three. They’ve done a pretty good job, but they still don’t have a lot of the premium UX that Grok Bot has. Honestly, on UX, both of the other model houses are far behind Grok Bot.

That leaves Claude third on UX, even with the better model.

Are they using what we’re using?

I have a theory about why this is happening. From my own experience, it feels like the Grok Bot engineers are actually using their own model. At the other places, I think they may be using variants, maybe a different version than what’s been released. So the people at OpenAI and behind Claude Code might be working with a different model or variant than what they’ve put out, and they’re used to it. They’ve tuned it and personalized it, and that’s not the same experience someone using the public version gets.

I don’t think they’re dogfooding enough. If they were, I think they’d notice these rough edges right away. To me, Claude’s permission theater is just unusable. At times you want to throw the kitchen sink at the monitor. It’s that frustrating.

This could change fast

Could all of this change in a week or two? Yeah, it could. But right now this is how it is, and I figured I’d talk about how Grok Bot has become a well-trained Western waiter with a great user experience.

When you’re using these things 14 to 15 hours a day, it makes a huge, noticeable difference, and I’m really appreciative of it.

That’s what I’ve got for you, hopefully it helps.

---

If you want agents that actually work for your business, go to agints.com/free, talk with Gini, and see where you can put agents to work.

Get the next one in your inbox → agints.substack.com