One Bot per Domain: Running My Personal Life on Grok Bot
How I split my personal admin across a fleet of Grok Bot agents—one per life domain, with a router in front—and what worked, where it broke, and which guardrails I'd keep if I started over.
Personal site note: views expressed here are my own and do not necessarily represent Red Hat or any other organization. This is not a sponsored post. Everything below is from my own use over the last few months; I’ve stripped out names, accounts, amounts, and anything else personal.
I spend my workdays telling customers to stop running everything on one big cluster with one big admin account. Split by blast radius, give each workload only the access it needs, automate the boring parts, and keep a human on the approve button for anything that can’t be rolled back.
A few months ago I realized my personal life was the opposite of that. One inbox, one calendar, a pile of portals, and me as the only operator. So I did what I’d tell a customer to do and broke it up. I now run a small fleet of Grok Bot agents, one per life domain, with a general-purpose bot in front that routes work to the right specialist.
This post covers what that looks like, what it actually got done, where it fell over, and what I’d do differently. It’s not a product review. Some of the most useful moments came from the bots telling me what they couldn’t do.
The architecture: a router and a set of specialists
The setup follows a hub-and-spoke model. The main bot, simply called “Grok Bot,” is where I type most things. It answers directly when it can, and when the work belongs to a domain, it hands it to that domain’s bot, either by messaging the bot or by having me talk to it directly.
Here’s the inventory, written the way I’d write it if it lived in a Git repo:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
fleet:
router:
name: Grok Bot
role: front door, cross-domain questions, hands off to specialists
specialists:
- name: Email # weekday morning inbox triage, drafts replies
- name: Calendar # Google Calendar reads and writes
- name: Bills # utility, phone, and service bills
- name: Finance # accounts, spending, investments (read-mostly)
- name: Home # house records, insurance, warranties, vendors
- name: Medical # appointments, portals, paperwork
- name: Tickets # event tickets, resale, conflict checks
- name: Music # taste analysis, show and festival picks
- name: Travel # flights, hotels, credits
- name: Buy/Sell # marketplaces, price checks, listings
connectors: [Gmail, Google Calendar, Google Drive, Google Docs, Plaid (read-only), X, GitHub]
secrets: vault # logins live there, not in chat
approvals: required # anything sent, posted, bought, or deleted
Why many bots instead of one? The same reasons you’d split workloads on a cluster:
- Context stays clean. The Home bot knows the house, the insurance history, and the vendors. It doesn’t need my concert history, and the Tickets bot doesn’t need my mortgage.
- Memory is scoped. Each bot keeps notes about its own domain, so a question like “is the house still under the builder warranty?” gets answered from records that bot already gathered, not from a cold start.
- Routines have an owner. Every scheduled job lives on the bot that owns the domain, so I know where to go when one needs to change.
What the specialists actually did
I’m only including things that happened. Details are generalized.
Email: the morning brief
The Email bot runs a weekday morning inbox check. It reads the overnight mail and sends me a short note: what needs a reply, what looks like a scam, which security notices describe activity I should verify, and anything with a deadline. Most mornings the answer is “nothing urgent,” and that’s the point. I stopped opening my inbox to find out whether something was on fire.
It also drafts replies when I ask. When a home cleaning service billed me well above its own written quote, I said “email them and ask why.” The bot pulled the quote and the receipt, wrote a reply on the original thread that laid out the gap and asked for an itemized breakdown and a refund, and showed me the draft. I hit send. It then watched for the response in the morning brief.
Bills: the portal migration trap
I asked the Bills bot to pay a utility bill. The old portal showed a zero balance, which would have been the end of it for me. The bot noticed that the bills had moved to a new billing system months earlier, opened the PDF, and found two earlier bills that had never been paid because they went to a portal I didn’t know existed. It registered the new account with a password generated in the vault, showed me the real total before charging anything, paid it after I said yes, and then set up AutoPay.
For the AutoPay step it needed bank details. It didn’t ask me to paste them in chat. It put up a secure form, the values went straight into the portal, and the bot said plainly that it never saw them. That’s the right pattern.
The same bot helped me push my phone carrier for a better rate. It broke down what I was actually paying for the plan versus a device payment, which made the comparison with cheaper carriers honest. The carrier said no to a match and offered a plan review instead. More on what happened next in the failure section.
Finance: mostly read-only and skeptical
The Finance bot links accounts through Plaid with read-only access, and it reads alert emails for daily notes on what moved. The most valuable thing it did was distrust its own data source. It pulled a full transaction export from a budgeting app and found the app had been listing the same card transactions twice under different card numbers. That inflated my spending by a lot and invented “duplicate subscriptions” that didn’t exist. It also flagged a cluster of large charges at one merchant and told me to verify them before paying the card.
Beyond read-only account monitoring, it also helped me with beneficiary paperwork across my investment accounts when I asked—the kind of task I’d been putting off for years.
Home: a binder that keeps itself up to date
The Home bot created a Google Doc called “Home Binder” and a Drive folder for the house, and it keeps both current: mortgage details, insurance policies, HOA documents with a plain-English summary of the rules, and warranty records. It tracked down a builder warranty transfer I’d done years ago and worked out exactly when each coverage tier had expired. It found that one policy I thought I had canceled was actually a separate flood policy that was still my only flood coverage. With my approval, it emailed my insurance agent for a declarations page, then watched for the reply on a weekday schedule.
It also talked me out of things. It said no to a cleaning service whose headline price turned out to be an auto-renewing membership with an early cancellation fee. It said a home warranty didn’t make sense for a newer house. And when I asked whether to keep a property as a rental, it built a multi-year model and laid out which inputs actually moved the answer.
Medical: portals and paperwork, not diagnosis
The Medical bot reads appointment and portal notices from email and puts appointments on my calendar with the prep notes. It also booked a specialist appointment through the office’s online scheduler.
Patient portals are where it hits sign-in walls. It opens the portal on its own desktop and hands the screen to me to sign in, then picks the work back up. It also drafted the cover letter for an application form and told me which fields I still had to fill in and sign myself.
Tickets, Music, and Travel: the fun stuff
- Tickets relists tickets I can’t use. It found the original order, uploaded the PDFs, set a price, and later dropped the price when I asked. For a festival wristband, it steered me to the festival’s official resale partner and pointed out that, under the festival’s terms, tickets resold anywhere else might not be honored. When a resale deadline was coming up, it set a check that runs a few times a day to look for the ticket email.
- Music read my SoundCloud likes (Spotify needed a sign-in on its desktop), summarized my taste, and ranked festivals and shows for a weekend I wanted to celebrate. When I said I didn’t want to see one of the headliners, it went back and found better options.
- Travel booked a flight on my airline account. It covered the whole fare with an existing flight credit, got the exit-row window seat I asked for, and showed me the $0 total before booking. It also pulled hotel options near a festival and noted the free cancellation deadline.
The router: cross-domain glue
The main bot handles the questions that span domains. I told it I had canceled two trips and asked it to verify. It sent a read-only subagent through my email, calendar, and transactions, and came back with each booking’s status. One “cancellation” was actually a ticket resale, and I still needed to deliver the ticket. One refund had come back as airline credit, not cash. It also found the auto-created flight events still on my calendar and deleted them when I asked, then searched again to confirm they were gone.
Where it broke, or nearly did
This is the section I’d want to read before trying any of this.
Sign-in walls are the main bottleneck. Bot protection on a budgeting app blocked the Finance bot’s sign-in outright. The fix was low-tech: I exported a CSV myself and dropped it in. eBay stopped the Buy/Sell bot from viewing a live auction listing more than once, so it fell back to reading eBay’s emails. Patient portals and streaming services needed me to sign in on the bot’s desktop. Plan for a human in the loop at login, every time.
Platform rules beat good intentions. The Bills bot tried to send my reply to the phone carrier’s support account in X Chat. X blocked it because encrypted chat only goes to accounts that follow you, and the carrier’s support account doesn’t follow me. The bot told me it didn’t go out, gave me the text to paste myself, and explained why. A Facebook group declined a resale post under its anti-scalping rule. The bot paused its hourly buyer-reply routine, made a redacted receipt screenshot, and resubmitted the post with my approval.
Inferred is not known. The Email bot drafted a message to an equipment rental company and included a detail about my situation that it had inferred but never actually confirmed. I approved the draft without noticing. Right after it went out, the bot flagged its own mistake, said what it actually knew versus what it had assumed, and offered to send a correction. I’m glad it owned the mistake. The lesson is that I still have to read drafts, not just skim them.
Proactive can be stale. The router sent me a helpful check-in tip for a flight that I had already canceled a week earlier. When I asked it to verify the cancellations, it found the cancellation and told me to ignore its own note. Proactive nudges are only as good as the last time the bot checked.
Bots should push back on conflicting instructions, and mine did. I told the Tickets bot to drop a price the next day if the tickets hadn’t sold. Later I told the router to drop it right away. The Tickets bot noticed the conflict, refused to guess, and asked the router to have me confirm one option. That’s the behavior you want from an operator.
Path A vs Path B: one bot or a fleet?
Path A, one general bot. Less setup, one place to look, and memory isn’t split. It gets noisy fast, and every routine competes for the same context. It’s fine if you mostly ask one-off questions.
Path B, a router plus specialists. More bots to keep track of, and work can fall between them. In exchange, each domain gets its own memory, its own routines, and a clear owner. It’s the right call if you want recurring work, like bill checks, inbox briefs, and renewal reminders, to run without you.
I went with Path B and would again. I’d just add specialists more slowly.
Takeaways
- Split by domain, not by task. A bot that owns “home” builds up context that a bot asked one-off house questions never will.
- Keep approvals on anything you can’t take back. Sends, posts, purchases, and deletes all went through me. The one time I rubber-stamped a draft is the one time a wrong detail went out.
- Credentials go in the vault and secure forms, never in chat. The bot used a secure form to set up a payment and reported that it never saw the bank details.
- Expect sign-in walls and design for the handoff. Exports, email alerts, and “your turn to sign in” screens are the real integration layer for personal finance and healthcare.
- Make bots show their sources and their doubts. The most valuable output wasn’t a task completed. It was “your budgeting app is double-counting” and “this cancellation is actually a resale.”
- Schedule recurring checks to save attention. Let routines handle repeat checks, and keep big one-off requests scoped.
- Prune the fleet. Every half-finished bot is one more place for work to stall.
If you work in platform engineering, none of this will surprise you. It’s least privilege, separation of concerns, and change approval, applied to the stuff that used to eat my weekends.