Trust, But Verify

Welcome. If you’re reading this, thank you — genuinely. Whether you found this site through the agentic AI writing or the Bible study, I’m glad you’re here, and I don’t take it for granted that anyone spends their time reading a stranger’s site.

Now let me tell you something that should change how you read everything else on it.

This field is younger than it looks

Every post here that touches agentic AI comes out of a system I actually run day to day — not a demo, not a weekend project. So let me answer the question I keep getting asked, and one I should’ve answered up front: how long has this actually been running?

The whole project started April 6, 2026. The dedicated host it runs on today came online eight days later, April 14. As I write this, that’s a little over four months.

I say that not to undersell what it does — it does a lot, more on that below — but because agentic AI as a category is moving fast enough that “four months old” is already a meaningfully different system than the one I started with. What’s true about it today may not be true in three more months, and what I write here reflects today. You can’t trust any of this 100%, and neither can I. That’s not a flaw to apologize for — it’s the actual state of the field right now, and pretending otherwise would be the dishonest version of this post.

What it actually does

To make that concrete instead of abstract, here’s a real list — not a highlight reel, the actual working set of things running day to day:

• Files, scans, and receipts into a self-hosted document archive (Paperless-ngx)
• Workflow automation for repetitive multi-step tasks (n8n)
• PDF generation and manipulation for reports and reference material (Stirling PDF)
• Home automation and smart-device control (Home Assistant)
• Media library management
• Email, calendar, and document work across a full office suite (Google Workspace)
• Video platform management — descriptions, captions, playlists — for a ministry’s YouTube channel
• DNS, hosting, and deployment for a couple of small websites (this one included)
• Audio transcription (speech to text) and voice synthesis (text to speech)
• AI image and video generation for design work
• A local model for the routine, low-stakes stuff, so the bigger model’s attention goes to judgment calls
• Browser automation for sites that don’t have a clean API
• All of it running on a small self-hosted homelab — a few machines in a closet, not a cloud bill

That’s the honest inventory. Some of it is genuinely useful. Some of it is me finding out where the edges are.

The septic tank

Here’s the story that actually convinced me to write this post.

My septic tank was overdue for a pump-out — the kind of task everyone puts off. I mentioned it once, in passing, the way you’d mention it to anyone. What happened next is worth walking through step by step.

It looked up the tank size in my files in Paperless-ngx that I self-host. It figured out roughly how urgent “overdue” actually was for a system that age. It found several local septic companies, reached out to each one by email asking for pricing and availability, and came back with real quotes from real companies. Then it made a call: one company was newer, smaller, hungrier for reviews and repeat business than the 40-year veteran down the road — and reasoned that newer companies price more aggressively to build a client base. It recommended that one, guessed it would land close to a hundred dollars cheaper than the next-best quote, and it was right.

I said go ahead on the company. It checked my calendar for a time I was free, confirmed that time with me, and then booked the appointment. Those two check-ins — approve the company, confirm the time — were the whole extent of my involvement. I never called, emailed, or negotiated with anyone myself.

A guy showed up at my house on the scheduled day and pumped out my septic tank. He had absolutely no idea that the entire chain of research, outreach, comparison, and scheduling that got him there had been orchestrated by an AI while I was lying on my couch.

The contractor who showed up at my door had no idea the whole thing was run by an AI while I was on the couch. I still had to say “go ahead” before any of it became real.

That last sentence matters more than the trick of it. The interesting part isn’t that it could do the research — it’s that even with all of that done, nothing actually happened until a person said so.

Checks and balances, on purpose

That’s the model I actually hold myself to, and the one I’d ask you to hold me to as a reader: research, comparison, and recommendation can be delegated. The decision that spends money, commits to a date, or acts on your behalf in the world stays with a human, every time. Not because the tool can’t be trusted with judgment — it clearly has some — but because “usually right” and “always right” are different claims, and only one of them is safe to build a habit around.

So take the septic story as what it is: a genuinely useful thing that happened, from a system that is four months old, in a field that is maybe five years old, being written about by someone who still checks the guy’s work before he leaves the driveway.

Read what’s here with that in mind. I will keep writing about what this thing actually does, good and bad — but always, always, do your own homework before you act on anything you find in these posts, mine very much included. The responsibility rests on you as to what and how you implement what you may learn here, and so do the consequences, so be careful.