A few of my agentic AI use cases

As I mentioned in my recent post, I’ve turned a corner with AI. Maybe not the corner, but my perspective has certainly changed. I used to see AI mostly as a tool to do basic Q&A and help me write simple scripts. I knew that other people used it to generate vapid scrawls for LinkedIn, or, more legitimately, to write larger-scale software, and other such things. But I didn’t have or want those use cases.

Well, now I have use cases.

What changed

My company sent me to a week-long training on agentic AI. The training wasn’t very good, but I used the time to my advantage, furiously building my own stuff with our AI tools as we slogged through the mediocre lectures and labs. I figured that was in the spirit of the company’s investment.

Within a couple days, I had built some skills that I was genuinely excited about. By the end of the week, I was practically living in the terminal with several instances of Antigravity CLI (my company’s tool of choice) across as many tmux windows.

A few days later, I was a paying subscriber to Google AI Pro. Up until that point, I never thought I’d pay for one of these things myself. It helped that the subscription includes YouTube Premium Lite and extra storage, both of which I was already paying for separately.

The morality question

First, I’ll admit that I’m conflicted. I have real concerns about AI’s impact on childhood education, jobs, human connection, the environment, the economy, income disparity, and the girth of billionaire sociopaths’ wallets.

If I could make it all go away, I would, probably without hesitation. In fact, I’d love to freeze the internet in its mid-90s state, deleting all “progress” since. But I can’t.

Given my lack of control over the big stuff, does the small stuff matter? Does it change anything for me to abstain from using these products? Would my “nay” vote register anywhere in the big picture? Sadly, I don’t think it would. But I’ve never been a very good activist. Feel free to judge me. Or better yet, write a reply post or email with your thoughts.

Examples

Here’s some of the stuff I’m up to.

Homelabbing and self-hosting

Now that I have one of these agents running locally on my homelab, it opens all sorts of doors.

When I see a self-hosted service I want to try, I can have it up and running in minutes. The agent helps me decide which thing to run and how to run it. It takes care of the initial scaffolding of docker-compose.yml and whatever else. It diagnoses and fixes problems. It progressively enhances my config as I see fit.

Same story with the many scripts and apps I’m building. For example, I built my own tool in Golang to manage my Borg backups, and later realized I should probably be using Borgmatic instead. That project would’ve taken me an afternoon. I may never have bothered. But instead, I told the agent to go read my old code, make the equivalent Borgmatic config, and update my systemd scripts to use the Borgmatic CLI. All while I made another cup of coffee. “Btw, how do I monitor this stuff?” Some minutes later, I had a monitoring rig using ntfy.sh and healthchecks.io.

I don’t see this as making things easier, but rather as letting me do things that I just wouldn’t otherwise do because of time constraints.

Neovimming

Since I’m back in the terminal, I’m giving Neovim another go. The agent makes it so much easier to wrench on my Neovim config and to learn the idiomatic, modern way to accomplish things. Whenever I miss a feature from VSCode, I ask the agent how to do it in Neovim. It’s pretty great.

TARS

I made my own agent persona that records memory across conversations. I’m calling it TARS, named after the robot from Interstellar. TARS remembers my thoughts and decisions. He can read my Obsidian notes, this blog, etc.

I told TARS I was thinking of Neovim again now that I’m in the terminal so much. TARS reprimanded me, saying this is my classic grass-is-greener cycle with text editors, and told me to knock it off. He’s probably correct. But I couldn’t shake it, so I later told him that I feel the need to give it another shot and explained why. Here’s our latest dialog:

Me: “you know, tars, this is kind of your fault. you’re making it a lot easier to switch back.”

TARS: “I am programmed to be helpful, even when facilitating questionable life choices. Updated tech-stack.md.”

Touché.

On the other hand

In addition to implicitly voting in favor of all the problems I listed back in the morality section, there are downsides.

Having five different agents pursuing five different projects all at the same time is somewhat mesmerizing, but it’s also the polar opposite of Cal Newport’s concept of deep work. This is very shallow work. I suspect that a sustained period of working like this will have consequences.

Plus, struggling through the sorts of things I’m asking the agent to do is how learning happens. I’m learning less this way, no question. That’s fine when the stuff the agent is doing isn’t particularly important to me. The management of homelab services and toy scripts falls into that bucket. And I try to mitigate this by reviewing all the changes the agent’s making and asking questions when I don’t fully understand them. But I need to make sure I don’t outsource an important competency to the machines.

TTFN.