what I’m actually using Grok Bot for...

I tweeted my prediction for the future of AI agents. Less than 24 hours later, xAI launched Grok Bot and matched it almost to a tee.

Creating a new agent is as simple as filling out a questionnaire.

Each Bot gets its own cloud computer. It can sign into apps, store files, use a browser and terminal, remember previous work and run scheduled routines.

You can also continue the same conversation from your desktop or iPhone.

One thing I love is how simple the setup is.

The menu and user interface are almost nothing. There’s a Plugins section and a Settings section. That’s pretty much it.

A plugin can package MCPs, which connect the agent to external tools, and Skills, which teach it how to complete specific tasks.

The complicated bits still exist underneath, but you don’t have to think about them.

My existing skills library imported with one click (thanks to my skills atlas).

I tested it out and created three different agents:

They’ve all been doing a surprisingly good job in their respective fields.

I had Gordon repurpose last week’s newsletter across X and LinkedIn and start replying to my YouTube comments.

He now does it automatically, and the outputs are extremely solid with virtually no training from me.

While testing, I also found out that the Bots can message one another.

I asked Alfred to sign an NDA and he delegated it to Florence, who handles my brand deals.

One little gripe is Gordon spent an hour trying to post an article to X. I could not inspect what tools it was calling, understand why it was stuck or stop the task…

That simplicity is useful when everything works. When something goes wrong, you are left watching an agent burn through usage without many ways to intervene.

xAI includes Grok Bot with qualifying plans, including the $200-per-month Cursor Ultra tier.

If you already pay for Cursor Ultra, Grok Bot comes as an extra. If you don’t, another $200 subscription is a harder sell.

lots of people have been quick to sh*t on grok bot because of this, but I think the $200 comparison needs more nuance.

The most cost-effective way to run Hermes at this level is through a Codex subscription, which is roughly $100–$200 a month anyway.

I personally can't see it replacing my Claude Code usage, but I think I'll use them for different things.

Claude Code will be the application I open when I want to build things, and Grokbot will be an agent that I use when I have built amazing skills and I want them running autonomously on schedules, such as:

It has earned that role in my stack (for now…)

watch my 15 min Grokbot live build here

The internet is furious that Anthropic plans to add invisible watermarks to Claude’s output.

I don’t like it either, but the reaction has got ahead of the facts.

Anthropic says future Claude models will hide a watermark inside anything they write. You won’t see it, but a detector will be able to find it.

None of the current Claude models use it yet. Anthropic hasn’t released the detector or explained how it works.

Treat it as a Claude-contact detector.

If you write an article and ask Claude to proofread it, the finished version may still carry the mark. That shows Claude touched the text. It does not prove Claude wrote it.

The same EU rules cover the wider industry. OpenAI has backed them, and Google already watermarks some Gemini text.

I don’t expect this to have a major impact on my life or business.

AI SEO and meta ad copy are the only areas I would watch. I also expect watermark-removal tools to appear quickly too.

My guess is that this becomes one more annoying hoop, rather than the end of AI-assisted publishing.

I teased the Skills Atlas a few newsletters ago.

It’s now fully finished, and this is one of the projects I’m proudest of building in recent months.

I recorded my second podcast episode with Greg Isenberg this week while running him through the setup.

The more I build this stuff, the clearer the direction becomes.

Agent setup is going to get simpler and simpler. Your folder structure and technical setup won’t matter as much as they do today.

I think the average user will eventually have a Grok Bot-style agent running whichever models are best, with all the configuration handled for them.

I’ve been looking 12 to 24 months ahead and trying to work out what will still matter.

Time and time again, the answer in my head is Skills.

Skills are basically instructions that teach an agent how you or your company likes a specific job done.

Agents will keep getting smarter and easier to use. You’ll still need to teach them your processes, preferences and standards.

The Skills Atlas solves three problems:

The usage view shows which Skills people use most, which have gone quiet and what each teammate is running.

It plugs into Claude Code, Cursor, Codex, Grok Bot and pretty much any other agent.

One click gives that agent access to your company’s full Skills library.

I’m giving all my community members access over the next week. I also made a little read-only preview so you can open it and click around yourself.

Try the Skills Atlas preview here.

I tend to plot agent harnesses on two axes: how customizable they are and how difficult they are to use.

My first guess put Grok Bot in the bottom-left. It looked extremely simple, but not very customizable.

Then one of my developer friends pointed out…

that each Bot has its own full computer, you can click into the file system and configure pretty much everything about the agent inside it, including the files, Skills, plugins, workflows and memory.

That changed where I put Grok Bot on the graph.

The main thing you cannot control is the model. It stayed near the simple end, but moved much higher on customizability, into the blue.

Hermes and OpenClaw can already do much of this. But they mean setting up a VPS, installing in the terminal, setting up ssh, configuring the models and tools, and maintaining the whole thing yourself.

Grok Bot gives you much of that flexibility behind a normal app.