🔥 Jadwal & Hasil: p/personal ai agent grok bot vs claude duet - Info Live Streaming 2026

Welcome to the monthly paid member Q&A for October 2026, where I answer your AI questions live on Substack.

This one was just me. Like last month, it had two sections. First, I answered member questions. Then I went through What’s New in AI, and this month was mostly about one kind of product: the always-on agent.

Anthropic launched Claude Duet. Notion gave its agents their own computer. Grok Bot kept shipping something new almost every week. OpenAI’s dots and Meta’s Muse are already out there. And halfway through the session, one of members asked the question I think a lot of you have: if you put these personal agents side by side, which one wins?

Before we start, here are some related posts you may want to read this month:

1. Claude Code Desktop Now Runs in the Cloud, in Parallel, and From Your Phone: the Claude Code updates I mention in What’s New.

2. I Never Had a Chief of Staff, So I Built One With OpenAI Dot and GrokBot: how I use dots and Grok Bot together.

3. I Built a Seven Grok Bot AI Team I Can Run From My Phone: the original Grok Bot setup behind most of the examples here.

4. How I Use Claude Code Routines to Run Work While My Laptop Is Closed: the path I suggest if you only use Claude.

5. Turn Your Website or Slides Into a Video With HyperFrames: how to install HyperFrames before you try the video workflows at the end.

FYI, I’m also writing a full Grok Bot guide for Thursday, covering everything from beginner setups to God Mode. If you have Grok Bot questions you want answered there, send them over in the comment section below.

What’s my take on always-on agents

Before we continue, here’s what I think about the always-on agents trend that we’ve been seeing more and more lately.

Grok Bot has the most features right now, and I like being able to run several bots with different jobs and personalities. But in my experience, Grok 4.7 model isn’t as strong on long or complex tasks. Some of my requests stop halfway, and I have to nudge it to keep going. So if we compare at the intelligence game, I’d expect Dots and Claude Duet to win, because they run on GPT‑6 Astra and Claude Opus 5.5.

So I don’t use Grok Bot for deep and complex work.

Right now, my Grok Bot handles simple, recurring tasks across apps. But what’s interesting is that it can also hand off coding work to a Cursor agent running any model I choose, like Opus 5.5. The rest of this post will show you how I use Grok Bot and what type of work I’ll still be doing with Claude/ChatGPT.

What you’ll take from this month’s Q&A

Here’s one Grok Bot workflow from the session that shows the kind of thing we’ll cover.

I get a lot of sponsorship inquiries and don’t have time to read them all. So when an email arrives, I forward it to the email address of Ben, my lead Bot. Ben opens a task for my sponsorship Bot, which researches the company and tells me whether the advertiser is relevant to me, whether I should reply, and whether I need to know more about them. If it’s not a fit, I tell it to reply with a polite no.

All I did was forward one email.

Here’s what’s in the rest of the post:

• A scheduled form-checking instruction for Shuba: she was asking a form-filling agent that finds changes, retrieves the right form, checks its answers, and flags conflicts.

• A computer-use prompt for Juha’s missing sound: she wanted to find out whether AI could fix the missing audio on her computer.

• My full answer to “which always-on agent?”: where Grok Bot wins, where it stalls, and a short selection rule for picking between Grok Bot, dots, Duet, and Claude Code.

• Three Grok Bot workflows I run now, plus the Cursor handoff: an X bookmark library that refreshes every morning, a command center that tells me which Bot needs me, forward-an-email sponsor triage, and how I hand website changes to a Cursor agent on Opus 5.5.

• The Claude-only path: One of Maker Labs’ members shared how he used Claude routines and webhooks to turn Claude into his personal agent that works in the cloud 24/7.

• Video with Opus 5.5: Riley Brown’s one-prompt Blender edit, a prompt that finds archival B-roll on the Internet Archive, Every’s article-to-video steps, and the new HyperFrames brand kits.

A few honest limits before you start. I don’t have Claude Duet access yet, so everything I say about it comes from Anthropic’s announcement. And somehow, about three‑quarters of the way through the video, I accidentally turned off screen sharing and didn’t notice at all. But no worries—I’m sharing every screenshot and all the links I mentioned during the session.

By the end, you’ll know which work to hand off to an always-on agent and which work to keep in Claude. You’ll also get a glimpse of how you can start building your Grok Bot with more advanced use cases.

Let’s dive in.

Shuba: An agent that fills forms behind a login, with no vendor API

The first question came from Shuba.

Shuba wants a form-filling agent that detects changes, retrieves the right form, checks its answers, and flags any conflicts. She gets stuck in three places: forms she wants to retrieve require a login, some forms have moved and now return a 404 error page, and third-party data comes from vendors that don’t offer an API.

Find the right way to retrieve external data

Here’s how I think to tackle this problem.

The first question is how the agent reaches the data. I’m assuming it needs to retrieve the data from external apps.

If the app gives you an API, connect it through an MCP or a CLI. That’s the cleanest path, because the agent reads the data directly and nothing depends on a web page.

If the vendor has no API, there’s no direct connection for the agent to use. The fallback is the browser. The agent opens the site in its built-in browser, signs in like you normally would, reads or downloads what it needs, and writes the result into a Google Sheet or a document.

On the call, I showed this in Claude. I asked it in chat to browse aimaker.substack.com through the built-in browser, and it opened my site and started reporting back. That’s the same approach Shuba needs.

Let a password manager handle the login

Next, you need to let your agent log in to the app you want to retrieve data from; you don’t want to share your passwords directly with your AI. This is where a password manager can fix that. The agent asks to sign in, you approve the request on your own device, and the password manager fills in the login on the page without the agent ever seeing the password.

There are two versions of this right now: