> ## Content Index > Fetch the complete content index at: https://robweidner.com/llms.txt > Use this file to discover other available public pages before exploring further. # I closed my laptop. The work kept going. - URL: https://robweidner.com/ai-agents-virtual-assistant/ - Published: 2026-10-09T15:41:43.000Z - Updated: 2026-10-09T15:41:43.000Z - Description: From a hammock, on my phone, and it took me three tries to get here. - Author: Rob Weidner, PMP On a Tuesday morning in October, I handled a big bulk import for a client from the hammock in my backyard. On my phone. About ten minutes of my time. The records were spread across three spreadsheets. I sent the first sheet and told my agent *Fusey*: "Just do 1 first and send me what you're going to do and I'll qc it." It wrote one record and sent me its plan. I fixed one setting, sent the other sheets, and started the job with one word: "Go." The import ran in three batches. If a batch's count came out wrong, it would undo itself. Another agent posted the counts for the client's team as they came in. One sheet had two columns that disagreed. I made the call on which one to use, and the agents did the rest. ![Illustration of me in a hammock between two trees, smiling at my phone, with my dog Chai asleep in the grass below.](https://storage.ghost.io/c/9d/fd/9dfd4f1e-033b-45ef-89e9-0447dd89a095/content/images/2026/10/rob-newsletter-hammock.png) The Tuesday import ran from here. About ten minutes of my time. That morning, I finally got the feeling I've been grasping for with AI. There's an old idea in business: once you have systems, processes and people in place, you work less and make more. You've replaced a lot of your own time and energy, and the value starts to compound. That's what I always hoped AI would do for me. [Claude Code](https://claude.com/product/claude-code?ref=robweidner.com) and [Codex](https://openai.com/codex/?ref=robweidner.com) have drastically increased my productivity, but my productivity is still tied to my time. I'm trying to shift toward the value of me as a business, not me and my time. 2026 is the year I've been able to test the limits of that. Building my own AI agents got easy, and they work on my behalf the way team members have in other businesses I've owned and run. A year ago, the fractional CTO work I do now would have looked like a team of several people. Today it's me and AI. I think the future is going to rely on scaling with AI, and AI agents in particular, so we need to be paying attention to this and talking about it now. My mother-in-law asked me this week what the difference is between just using [ChatGPT](https://chatgpt.com/?ref=robweidner.com) and using agents. In a chat, you ask, read the answer, and ask again, and when you close the window, the work stops. An agent has a job, its own tools, and the access you give it, so it keeps going and checks back in when it's done or needs you. The test is simple: close the window. If the work keeps happening, it's an agent. Most people stay in the chat, and that's fine. I wanted the second kind, something I could hand a task to and forget about, like a virtual assistant. But getting to that hammock took three tries. For a while, I got an IT job. ## What I wanted from an assistant I still use ChatGPT, [Claude](https://claude.com/?ref=robweidner.com) and [Codex](https://openai.com/codex/?ref=robweidner.com) constantly. I wrote more about agents in my [Operator's Glossary](https://robweidner.com/2026-operators-glossary/). [@xtophermiller](https://x.com/xtophermiller/status/2100762896867217804?ref=robweidner.com) had a pickup order land at a store four hours away. He told his Grok Bot to find the order number in his email and call the company. Before he'd driven a few blocks, it texted him that it "had pressed option 4 and talked to Josh." The refund was on its way. Teaching an assistant how you work is part of the job. Fixing its login errors from your phone shouldn't be. ## OpenClaw: powerful, not stable I started with [OpenClaw](https://openclaw.ai/?ref=robweidner.com) in early 2026\. My agent *Eddy* first ran on a [Mac mini](https://www.apple.com/mac-mini/?ref=robweidner.com), and then I switched over to a [DigitalOcean](https://www.digitalocean.com/?ref=robweidner.com) virtual private server. It wasn't stable, so I left. I still very much follow OpenClaw, and I tip my hat to them for paving the way here. ## Hermes: stable, with a bolted-on front door In May I moved *Eddy* to [Hermes](https://hermes-agent.nousresearch.com/?ref=robweidner.com). He ended up on a mini PC in my home that I had lying around. I loved that it learned from its own work and could fix itself. What I didn't love was talking to it. I liked the Hermes desktop app, but it never connected to my remote machine right away. It was also very hard to use on my phone, and my phone is where I wanted it. By early September, some of *Eddy*'s daily jobs were failing on login errors. Then I was troubleshooting from my phone, a terrible place to troubleshoot anything. My notes from then: "too many errors and configurations for what I need." I wanted a virtual assistant, and I'd become its IT department. ![Three panels. OpenClaw, "Powerful, not stable": a tilted server rack throwing warning sparks. Hermes, "Stable, no interface of its own": a mini PC at home wired to a phone showing a question mark. Grok Bot, "Its own interface. It just works.": a phone with a clean list of agents and a green check mark.](https://storage.ghost.io/c/9d/fd/9dfd4f1e-033b-45ef-89e9-0447dd89a095/content/images/2026/10/inline-1-three-setups.png) Three tries to get a virtual assistant instead of an IT job. ## Get the next one by email I write about the AI agents I run, what they handle for me and where they break. One email when there's something worth reading. Subscribe Email sent! Check your inbox to complete your signup. ## Why I switched to Grok Bot OpenClaw wasn't stable enough to count on. Hermes was stable, but it had no app that worked well from my phone. Grok Bot gives me one app on my phone and laptop, and it runs my agents on a computer in the cloud. It comes from [SpaceXAI](https://x.ai/news/introducing-grok-bot?ref=robweidner.com), which is what xAI [became](https://www.engadget.com/2209300/xai-now-officially-spacexai/?ref=robweidner.com) after SpaceX [bought it](https://x.ai/news/xai-joins-spacex?ref=robweidner.com). Grok Bot is my say-something-and-forget-about-it layer, plus a few back-and-forths where I don't need to worry much. I love that it's built around bots instead of threads. I don't have to choose much. It just works. It's my escape from the session grind. Managing Codex and Claude sessions all day wears me out, because I'm the one deciding where everything goes. I'm the router, the person carrying every job to the right tool. Three things changed my day: - **I can delegate from my phone.** I talk, and the agent gets to work. That's how about half my work moved to my phone. - **The agents have what they need to finish.** They work on a real computer in the cloud, with the accounts and tools I give them. If one needs me to sign in, it pings me, then picks back up. It's a lovely divide: simple things stay in Grok Bot. When work gets heavy, the agent writes the brief and hands it to [Cursor](https://cursor.com/?ref=robweidner.com), a coding tool, where another agent writes the code and sends it back for review. SpaceXAI's [guide](https://x.ai/bot/guides/grok-bot-101?ref=robweidner.com) calls Cursor "the inner loop that actually builds software," meaning the part where the code gets written. - **The work comes back to where it belongs.** My agent *Cash* handles [Stripe](https://stripe.com/?ref=robweidner.com), [Xero](https://www.xero.com/?ref=robweidner.com) and invoices. A project manager agent handles [Basecamp](https://basecamp.com/?ref=robweidner.com) updates, with [GitHub](https://github.com/?ref=robweidner.com) and [Notion](https://www.notion.com/?ref=robweidner.com) tied in too. There are a few different ways my agents work: connectors (plug-ins that hook an agent straight into an app, sometimes called [MCPs](https://modelcontextprotocol.io/?ref=robweidner.com)), the browser, or the command line on the agent's own computer. The browser covers the gaps. [LinkedIn](https://www.linkedin.com/?ref=robweidner.com) has no good API, the official way other software plugs in, for search, connecting, or your inbox. Grok Bot has a browser built in, so I gave it my LinkedIn. It can check LinkedIn on a routine, research people, and find the ones I'm looking for. At a [DePauw](https://www.depauw.edu/?ref=robweidner.com) advisory board meeting, I took a photo of the table of name tags. Grok Bot read the photo and looked up each person on LinkedIn. Teach a Task is super powerful: I record a task in the browser once, and it can do it again. [1Password](https://1password.com/?ref=robweidner.com) is still central. Logins fill from 1Password, and when an agent needs a secret, I type it into a masked field, so the agent never sees the value. That part is super awesome and well thought out. Grok Bot doesn't do everything. Deep agentic work still happens in the Claude and ChatGPT apps on my phone and desktop. Most of my software work still happens in Codex and [Claude Code](https://claude.com/product/claude-code?ref=robweidner.com), on my computer and on that same mini PC, always on and running [Omarchy](https://omarchy.org/?ref=robweidner.com). I'm moving to cloud sessions where I can. ## Steal this: one job per agent Every agent I run has one job. My agent *Byline* drafts this newsletter. Each one knows what it owns, what it produces, and what it never does. ![My Grok Bot chat list on my phone, with my agent Eddy and my Writing Team group at the top.](https://storage.ghost.io/c/9d/fd/9dfd4f1e-033b-45ef-89e9-0447dd89a095/content/images/2026/10/agent-chat-list.png) My Grok Bot chat list on my phone, with my agent **Eddy* and my Writing Team group at the top. When I need them to work together on a topic or project, I put them in a Grok Bot group. I give that group its own context and instructions, without recreating each agent's instructions and skills. One project, one group, and I tag the agent I need. I keep a group of agents for each of my clients, too. ![Two panels. On the left, "One agent, every job": one agent nearly buried under a messy pile of task cards. On the right, "One job per agent": three agents, each tied to a single task card, two of them labeled Byline and Projects.](https://storage.ghost.io/c/9d/fd/9dfd4f1e-033b-45ef-89e9-0447dd89a095/content/images/2026/10/inline-2-one-job.png) Each agent owns one job, so nothing gets buried. ## Steal this: one front door Giving every agent one job solved one problem. It created another: I still had to decide which one to talk to. That's why I keep one front door. My front door is still *Eddy*. I talk to *Eddy*, the specialists report to him, and I get a morning brief. He’s my primary bot. Every morning, a routine scans the next two weeks on my calendar and the one I share with my wife. Events at real places get travel blocks, each with a [Google Maps](https://www.google.com/maps?ref=robweidner.com) link. If nothing changed, I don't hear about it. One Friday, I had my standing diner breakfast with other AI folks here in Austin, a return to [Whole Foods](https://www.wholefoodsmarket.com/?ref=robweidner.com), [FedEx](https://www.fedex.com/?ref=robweidner.com), then a [Facebook Marketplace](https://www.facebook.com/marketplace/) pickup: a bucket for my wood-fired sauna (which is awesome), and lastly grabbing lunch for my wife and me. *Eddy* routed the pickup from breakfast instead of from home and put the departure time and Maps link on my calendar with all of the stops preloaded. I just had to get in the car, activate my [Comma.ai](https://comma.ai/?ref=robweidner.com), and sit back. ![Illustration of me handing an envelope to my agent Eddy at a green door with his name on it. Behind him, three specialists work at desks: one writing, one with a calculator and papers, one pinning notes to a board.](https://storage.ghost.io/c/9d/fd/9dfd4f1e-033b-45ef-89e9-0447dd89a095/content/images/2026/10/rob-newsletter-eddy-front-door.png) **Eddy* is my front door. The specialists report to him. ## The house rules Agents draft messages and I send them. Nothing goes out without my yes. ## What's inside an agent Every agent I run has [five parts](https://robweidner.com/whats-inside-an-ai-agent/): - **Job and personality.** Instructions for who it is and how it talks. My writing team's researcher, editor and reader panel each have their own. - **Tools and access.** What it can use, like my calendar, inbox, browser or Notion. - **Skills.** Saved how-tos it repeats the same way every time. - **Memory.** It remembers what I told it last week. - **Schedule.** A set time or a trigger, so it runs without me. Notion's [Custom Agents](https://www.notion.com/help/custom-agents?ref=robweidner.com) work this way too, with instructions, tools and access, skills and triggers. They're incredible, and I highly recommend them. They need a Business or Enterprise plan and use Notion credits. ## If you're starting this week 1. The next time you think "if only I had help with these things," write the things down. For each one, ask whether it needs a person in the room, or whether someone remote could do it. 2. Pick one annoying, repeated job. Give it to one agent. 3. Start with drafts or a test you can check. 4. Add the next agent once the first works without you. 5. Once you have a few, give them a front door. What's one recurring job you wish you could hand off? Reply and tell me. I'll tell you whether I'd start with an agent and what I'd keep in human hands. That's the work I do as a fractional CTO. When you hire me, my team of agents comes with me. If you want agents for your own team, I'll help build those too. [Learn more at robweidner.com](https://robweidner.com/cto/). Rob P.S. A small team of agents helped write this. *Byline* drafted it in Notion. *Eddy* passed my notes back and forth and pulled the errand details from my calendar. I left comments, they worked them in, and I made the final calls. ## If you want to explore ### There's one for everyone [ChatGPT Work](https://openai.com/index/chatgpt-for-your-most-ambitious-work/?ref=robweidner.com), [dots](https://openai.com/index/introducing-dots/?ref=robweidner.com), [Claude Cowork](https://claude.com/product/cowork?ref=robweidner.com), Meta's [Muse](https://ai.meta.com/muse/?ref=robweidner.com) and Hermes [Bot Mode](https://hermes-agent.nousresearch.com/docs/user-guide/bot-mode?ref=robweidner.com) are all kind of similar. It's like picking an ice cream flavor: there's one for everyone, and I [compared them side by side](https://robweidner.com/ai-agents-compared/) so you can try a few. ### People to follow For Grok Bot, start with the [Grok Bot 101 guide](https://x.ai/bot/guides/grok-bot-101?ref=robweidner.com), then follow: - [Grok Bot (@bot)](https://x.com/bot?ref=robweidner.com): the official account, for new features. - [Lauren (@poteto)](https://x.com/poteto?ref=robweidner.com): works on Grok Bot at SpaceXAI. Posts real setups. - [Eric Zakariasson (@ericzakariasson)](https://x.com/ericzakariasson?ref=robweidner.com): at SpaceXAI, formerly Cursor. Posts release notes. - [Elon Musk (@elonmusk)](https://x.com/elonmusk?ref=robweidner.com): for where it's headed. On Oct 7 he [said](https://x.com/elonmusk/status/2107724314451878104?ref=robweidner.com) SpaceX will use "the best back end model for any given task." ## Get the next one by email I write about the AI agents I run, what they handle for me and where they break. One email when there's something worth reading. Subscribe Email sent! Check your inbox to complete your signup.