Big Machine

Transmission ·7 min

A year later, and the bots are still just bots

Rob Conery

Back in January of last year I was seated in one of my favorite little pubs (Off the Wall) listening to a Delta Blues singer go deep on Robert Johnson. His slide technique was transcendent, and his voice so full of soul that I felt like I was in that tiny club just outside the Crossroads.

I had no idea this man would be there that night. I had come to do some work with my new OpenClaw setup, wanting to see how it would work on my phone while I was away from my machine. I had it look through my email and then held my breath as it reached out to Stripe to check my sales… the security warnings were everywhere, and I knew that I was probably doing something that wouldn’t end well.

But that’s me, and that’s what I do. I jump first and tell you how it felt when I landed. I jump again, and then again, and a few times after that, each time adjusting the landing or maybe even trying to do a flip midair. My job is to let you know what I did to make it hurt, or, hopefully, what was the most fun.

Sitting in that pub that night was the most fun I’d had with AI up to that point. It organized my Obsidian notes for me, triaged a support email, and sent a buddy of mine an email, telling him a dumb story that was completely made up.

I remember thinking: this is a wild, wild time, and it’s only going to get wilder.

Note: In this post I share my experience over the last year with various implementations of AI “bots”, and trying to get them to help me code. I created a video about this, too, which you can view below.

On Again, Off Again, On Again

Since that time, the coding tools caught up with their capabilities. Claude introduced Cowork, Perplexity dropped “Computer” (just oozing with inspiration, that one), and Codex did something similar. Scheduled routines came after that, and I remember thinking one day:

It’s so much easier and probably a lot more secure if I just kick up a mess of MCP servers for myself, run scheduled tasks, and get this thing off my machine.

Sad stories were, of course, running rampant. Someone dropped prod and blamed OpenClaw (or maybe it was Hermes), another person lost their emails (lucky them), etc. It’s probably smarter to be more careful, so that’s what I did.

Instead of giving blanket access to Stripe, my email, and other services, I decided to create my own MCP servers that gave explicit tools and access to the things I wanted Claude to know about. I also created MCP servers for companies that were a bit slow on the draw, including ConvertKit, Trello, and Notion. They had decent APIs, so creating an MCP for them was no problem.

By that time, Claude introduced Projects for Cowork, which allowed you to attach context and instructions, as well as working from a local directory. This was perfect, as I could attach my Obsidian vault and have Claude run daily scheduled routines that checked for support emails, rolled up my sales for me, adjusted my daily tasks, and blocked my calendar.

It worked fine, until it didn’t. I wish I had a screenshot of this, but the scheduled tasks would routinely fail with “the [THING] connector isn’t part of this chat, so I couldn’t update your notes” where “THING” might have been Trello, TickTick, Notion, or some other service.

I still don’t know what that meant, but basically it was a shot in the dark whether I would actually have anything to show for the scheduled tasks. I stumbled along this way, and then ditched the scheduled tasks completely, running things by hand when I needed them.

Just like I did a year ago.

We’ve come so far… yet not at all…

And Then GrokBot Came Along

A few months ago I decided to try GrokBot because I was seeing a lot of chatter around it. I’m not really a fan of Grok, and if I had to desribe why, I guess I would say that “it’s just ‘mid’”. Not bad, not great… just… OK.

snap53

You quickly realize that the trick is to not ask it to do too much for you. It’s good at small tasks with isolated constraints, such as “find me a carry on suitcase suitable for international travel that is mentioned the most in YouTube travel videos, on X, and is highly rated. Keep the price under $400”.

I’m a bag person, what can I say.

Grok found me a bag, which I almost bought (the TravelPro Carry On Elite but with Black/Neon trim) but I ended up with the Belroy Transit, which seems kind of small compared to my Monos…. anyway the point is this: Grok is OK, it’s just not great.

I gave it a solid two months, and I have to say the team (Matt Palmer, Lauren Tan, and gang) are very open and transparent. I love Matt’s YouTube updates, and I love the seamless way GrokBot works on my computer, in the cloud, in my car, and on my phone. It really is the best experience of them all.

It’s just… mid. Not something I would trust with deep work, like coding. I know there are others out there who use Cursor and like the models… it’s just not for me.

Not Far Behind: Claude Duet

I got in to Duet a few weeks ago and so far I really like it. Same kind of “Anthropic Confusion”, however, when it comes to knowing how and where your model is doing its thing.

snap54

In short: Duet runs in the cloud, which means it can’t directly access your computer. Sort of. You can attach files and those files get popped onto it’s cloud “VM” (Duet has its own computer, like GrokBot). Other that than, your sessions are always in the cloud.

But.

Duet can see your remote coding sessions in Claude Code, and you can also work with Cowork tasks. It’s confusing as hell, so I asked Claude to explain it to me:

Cowork is Claude's way of doing longer jobs for you. 
Each job is a task, and you can start one from the Claude app.

- Where it runs: each task runs in the cloud.
- Linked to a computer: a task can be linked to one of your machines, 
  and then it can use files and tools on that machine.
- How I fit in: I can read your tasks and send messages to some of them

If you have Remote Control turned on for Claude Code, Duet can send that remote session a message on your behalf. Why you wouldn’t just flip over to Claude (the app) and access your remote sessions yourself… well… I don’t know.

The subagent story is interesting too. There’s an “Everything” channel, which, like GrokBot, is your coordinator. GrokBot works with the idea of a staff being coordinated by a main bot. You talk to the main bot, it delegates work. I like that.

Duet is the same: you post something to “Everything”, which is basically the root, and it gets routed to a channel as needed. Those channels have access to Everything context, but they are separate sessions entirely, with shared context if that makes sense.

Where it gets interesting is the actual work: that’s done in yet another step down in Tasks. Each channel can spawn one or more threads, or “tasks”, that have a status like “Up Next”, “Scheduled”, “Active”, or “Resolved”. You can think of these as subagents with their own isolated context and prompt, that do not share context with their parent channel.

snap55

OpenBot: A Fresh New Face

I really like OpenBot and I can see it quickly becoming a favorite for a lot of people. Lots of neat interface tricks, cute, funny little avatars, and a very polished presentation. It has a sidebar that will ping you if you work in other apps, and there’s also a “Dynamic Island” that does the same, which feels like the MacOS notch in my screen (no thank you).

snap50

It also has a visual workflow builder, which I really like. It feels like n8n, and allows you to diagram things you want your bots to do for you, based on some kind of trigger:

snap51

Scheduled tasks also have a calendar interface, which I think is a great touch:

snap52

I would be using this thing full time if one of the following were true:

  • I was an OpenAI/ChatGPT maxxer.

  • I used OpenRouter or a local model.

  • I didn’t code very much, so my $200 API credit for Claude Code would work here.

This is the problem with all of the bots that use Claude, that aren’t Anthropic properties: you have to use your Anthropic API credits.

I hate that Anthropic does this. Deeply. Their models are simply the best for what I do, or maybe I should say that I’ve learned them the most.

OpenBot is also the only bot that I’ve had true success with when it comes to setting up coding tasks. It will automatically load your AGENTS.md and any skills you have in .claude because it is the main Claude session when you’re working in a code directory.

It threads conversations elegantly and the phone app is seamless and simple to use. I am SERIOUSLY impressed with this thing, so much so that I might give OpenAI another shot at writing code for me.

So: can you get bots to write code for you? Should you?

Yes, you can - but no, I won’t. One day I’m hoping to have a bot run things from an internal perspective: finding and fixing bugs proactively, triaging support issues, enhancing UX, and updating content that I want to add or remove.

I could do this now, I think, but honestly I don’t think Claude Duet is up for it. It’s also completely unneeded. I would rather open up a few sessions remotely in Claude Code, each devoted to that task. I would then schedule runs throughout the day to ensure things are working, customers are happy, and content is current. I might then, once a week, have a product designer bot recommend some fun changes and updates.

That’s a lot of work, and it’s a lot to maintain in terms of my attention span and overall “fiddling index”. I like to keep that to a minimum.

This is what bots are good for: they’re proactive little helpers that are supposed to act in your best interest. You give them a paragraph or two, and they figure the rest out on their own with your guidance.

For me, for now, I’ll stick with Claude Code for coding and my little bots for everything else. Travel, marketing ideas, content ideas, buying me suitcases… I’m using both GrokBot and Duet to see which one ultimately settles in my brain and hopefully I’ll be able to use one to both things well.

If this interests you, here’s a video! Also: I would love to hear what you’re doing, so shoot me a reply.

12-bots-thumb-1280