Dot is helpful, but not clever enough yet

Dot helped me work out what to do with my airline miles and Amex points. Once I helped it sign in, I could leave it checking the options while I did something else. That saved me time. The frustrating part is how often I still have to work out whether it is progressing or stuck.

It kind of works like Grok Bot: you message an assistant, give it access to the tools it needs, and let it work on a computer. From my phone, I can open Dot’s computer and see its desktop. It can also work on my Mac when I give it access.

I can see the appeal. I’m also finding that trusting an assistant to act involves more than trusting its answers.

The useful work happened before booking

I had several questions tangled together. What miles did I already have? Could I use them for this trip? Would buying more miles make sense compared with paying for the flight?

Dot checked balances and expiry dates across airline accounts. It compared paying cash, using United miles, topping up miles, and booking through Amex Travel. Having those options in one conversation helped me understand what I could do with my points.

Getting there involved cooperation. I signed in locally and checked things with Dot when access needed my attention. Once it could continue, I could get on with something else. I didn’t have to stay with it through every comparison.

I eventually chose a cash fare on United through Amex Travel and completed the payment myself. Dot then helped carry the booking into Tripsy, the trip website, and the plan shared with friends. The bigger win was understanding the options before I spent the money.

A familiar name changes the trust question

Dot and Grok Bot feel like more professionally packaged versions of an idea already visible in Hermes Agent and OpenClaw: a personal agent that remembers context, uses tools, and keeps working between messages. Those projects already support persistent assistants. The interesting shift is seeing that idea offered as a product where more of the setup and upkeep belongs to a provider.

I’ve been hesitant about trusting something like Hermes with that much access. Does putting an OpenAI or Grok name on the product make me more comfortable? A familiar company may make it easier to try. It still leaves me deciding which accounts to connect and how much authority to give it.

The open-source projects address those questions too. Hermes documents approval and isolation controls. OpenClaw documents tool restrictions, security audits, and incident response. Grok Bot’s own approval guidance says its model-based review should accompany restricted access and explicit approval boundaries. A polished interface doesn’t settle any of that.

What I want to understand is what the assistant can access, when it needs permission, and how I can check or stop what it is doing. If it makes a mistake, I need to know what changed and what can still be recovered. Some actions can’t simply be undone.

The ordinary failures matter too

Dot is too slow for some of the things I ask it to do. Sometimes using Codex with computer use feels faster. The unclear feedback makes the waiting worse: I can’t always tell whether it is working, waiting for me, or unable to continue.

When it needs a login, it should name the browser and computer. I’ve had to establish whether it means Brave on my Mac or its remote browser, and repeat instructions about where the work should happen. Helping it sign in is part of keeping control. Figuring out where it needs me is extra work.

File downloads are uneven. Dot downloaded 61 PDFs from Robinhood, but I had to download my Progressive policy PDF myself. I can’t assume a basic task will work because a similar one worked before.

Away from Wi-Fi, the connection has been unreliable for me. Talking to Dot in the car in San Francisco, I’ve had calls cut off as reception went in and out. I want the app to keep a typed message or voice request, mark it as queued, and send it when the internet returns. Then tell me when it has been received and when work starts. I shouldn’t have to remember to ask again.

That is where this class of product gets interesting to me. I’m handing over a job that may outlast the conversation. The assistant has to keep track of it, respect the limits I set, and make the outcome clear. The miles research showed the value of that arrangement. For now, too much of keeping it on track still falls to me.