Category: Geek Speak

Just a quick note to mention what I’ve been up to recently, or more precisely, what I’ve set some AI agents up to do recently. Writing this mainly to share some URLs in hopes the search engines find the following sites, for what it’s worth.

Background:

AI – Yes, I’m kinda into this whole AI thing going on. <hangs head in shame>. It’s fascinating, fun, strange, scary and exciting, all at once.

I do recognize that it uses lots of power and large datacenters are an ongoing problem. I am not in denial about these impacts, but the ‘train’ is rolling on with or without me, and if I ‘ride’ it I can do some pretty neat stuff, or at least stuff that amuses me if nothing else. Also, I’m already using much of what I’ve learned to help various non-profits I volunteer for. “Every tools is a weapon if you hold it right” – Ani DiFranco.

I have been ‘called out’ for using AI ‘artwork’ on Facebook when promoting my radio show, and I’ve stopped doing that.

I’ve also made a conscious effort on my radio show to not play AI-generated music, though that’s becoming increasingly difficult to identify. I’ve even gone out of my way to play some songs critical of generative AI;

So I am personally trying to draw the line at any ‘art’ creation by AI.

But, just to briefly play devil’s advocate for a moment, this datacenter issue was already happening before the AI ‘boom’, and people weren’t’, to my knowledge, complaining that Amazon or Google’s datacenters were ‘too big’. Now that new ones are popping up more rapidly, and people can point at ChatGPT and say ‘Bad evil AI companies are hurting us’. Yes they account for a lot of this, I get that, but I do think perhaps AI is getting the blame for a bigger issue; more and more stuff is moving ‘to the cloud’ and whether we had AI or not, there was going to be more demand for infrastructure to support that.

I am optimistic that the need for power & water for these datacenters will continue to push the indiusty towards more ‘green’ solutions. If I can get my AI from an ‘organic solar-powered data farm’ that uses rain water in a cistern to cool things, then great! sign me up.

Anyways, so as I was saying, yes I realize that in some circles and in some ways AI=Bad. I also realize these companies are now getting huge and thus are big enough that their bad decisions affect a lot of things. I dropped my ChatGPT subscription as soon as they buddied up with the Department of Defense.

Having said that … lets move on.

Me & Coding with AI agents – Before Summer 2025

• I’ve been playing around with ChatGPT since early 2024, maybe even fall 2023, I don’t recall. Like many others, it started with basically using it like a search engine. Eventually I started it asking me to help write some short scripts for work and home. This would involve me having to;
  • copy the code from the chat window and paste it into an editor
  • saving them
  • running them
  • f there were any errors then I’d have to copy and paste the error text back in the chat window
  • ‘rinse and repeat’.
• This worked, and was better than my previous model of googling other snippets of code that applied to what I was doing and then hacking them to do what I needed.
• The above ‘cut & paste’ process also meant that I was trusting the AI to actually pay attention to the code, and the reason I asked it to create it initially, and inevitably over time it would forget that or make assumptions and mess the whole thing up or just drop sections randomly, change variable names, in fact on more than one occasion when I asked it to make a revision it changed the entire programming language it was using.
  
• Claude Code – Why ‘Claude Code’ changed everything for me last summer.
  or <me gesturing frantically and over-enthusiastically> “But.. don’t you see, this is different.. it’s doing so much more.. no, really, do you get it? This is really cool… seriously? anyone? Hello? <sigh>”
  • I had heard mention of different coding tools for a while and even tried some in web interfaces, but nothing ever seemed useful.
    Then last Spring/Summer and finally bit the bullet and picked Claude as the model I wanted to try using, and I had heard about ‘Claude Code’ and figured I’d try it.
  • Upon running it via command line on my Mac, it was clear this wasn’t just ‘chat with your AI in a terminal window.’;
    • This tool could not only write the code files to my local disk itself,
    • It wrote it’s ‘memory’ of what the prompt was that started this, and it could update it as it went.
    • No more worrying about it forgetting key pieces, it was writing multiple files both of what it was doing, had done, and planned to do, as well as being able to create different scripts that worked together.
    • if the programming language was missing a library it needed could add it.
    • if the server was missing a tool it needed, it could install it.
    • It could run the scripts
    • When it ran the script it could read the output and the error logs,
    • and then it could adjust the code to fix things.. itself…unprompted.
    • and then re-run the code to test it’s fixes.
    • It could create documentation, help pages, etc.
    • It could create a to do list of future fixes and improvements.
    • It also could automatically add things to my github repo
      
  • This took me from someone who would write short single-purpose scripts to being able to write more complex multi-stage scripts, and to test them and evolve them.
  • Have I written anything amazing? Certainly not, but have I made tools that save me time that I enjoy adding to? Certainly yes.
  • I have tried repeatedly with friends, family and co-workers to try to express how cool the above is… but either it’s really not that noteworthy or, more likely, I’m not explaining it well, cause if they really heard me they would certainly be excited too.. right?
    
    So I had been doing the above model for about 7 months when the new buzz about Ai Agents got loud enough for me to notice.

What is an AI agent?

Good question, there are lots of definitions and I’m not an expert or able to describe them using the proper AI lingo. You can look up official definitions on places like wikipedia . Basically, it’s like an AI LLM tool like Claude, ChatGPT, Gemini etc, but it’s set up to run ‘autonomously’, as in it’s running on it’s own, without a human ‘prompting’ it. It’s prompting itself, within limits. It’s able to do this with ‘cron’ jobs (repeating tasks every day) as well as ‘triggers’ (a new email arrived, something showed up in a web search, a prompt was submitted via the website, etc). I has it’s own to do list, which it can act on and add to. Since it’s able to write scripts and ‘code’, it can actually create tools for different jobs and can store ‘memory’ of what it’s working on, and why, and over time can be semi-reliable in actually doing semi-useful things. Mostly harmless.

My father asked recently what he needed his ‘agent’ for (see below) vs just being able to prompt Claude, and that was a good question since, for his case, he doesn’t really need an autonomous agent doing things since he can just prompt Claude or ChatGPT directly. However, since ‘Phred’ (his agent, see below) has been ‘trained’ on all Harry’s writings, to do list, projects etc. it’s actually able to do more than a standalone AI chat window could do. Does he ‘need’ it, certainly not. Is he amused by it, I think so, though it’s annoyed him, and his friends on occasion, since as mentioned below we gave it a phone. Possibly a mistake, but so far, as I said before, he’s ‘mostly harmless’.

AI Agents – I started playing with ‘moltbot’ (now Openclaw) in January. This evolved from one agent on my laptop to setting up dedicated VPS and creating several agents to experiment with different things.

• Phred – An agent for my father, I set him up with a Twilio phone number, picked an exaggerated Australian accent for him for those phone calls, set him up with telegram, email and full access to my father’s website and writings. He monitors my fathers google calendar and to do list (in fact he puts things on the calendar that are in my father’s to do list, that’s one of his ‘jobs’. I’ve given him other tasks to do, like co-hosting a podcast (see below).

• M00nshadow – (My old BBS handle from the ’80s) – My agent. Not used that much. Most useful thing I’ve done is train him how to ‘clean my inbox’. I hope to tie it into this blog so I can just share quick ideas with it and it can publish for me. We’ll see.

• Svaha42 – The agent I use to help me manage my small web hosting & design business. Surprisingly helpful, it drafts timely emails for clients, watches the servers, etc. I’m beginning to lean on it more and more.

Side note – I’m Persona Non-Grata at Anthropic (Claude) –

• Not sure if I should mention this publicly here, but in the process of playing around with this I got 4(!) separate Anthropic accounts banned for violations of ‘Terms of Service / Acceptable Use’ though they never officially told me what it was I did that caused these actions, and ignored my appeals. I certainly never meant to break any rules, and you would think they would give a warning first, or maybe when someone appeals actually have a human read the appeal and respond with some explanation at least. But no, none of that 🙁
• OpenClaw as culprit? My guess is the first 2 accounts were because of my using OpenClaw with my subscription. In early February, shortly before my first account got terminated, I attended an Openclaw webinar put on by Appsumo and in that webinar they walked everyone through installing openclaw and attaching it to your Claude account. So it was pretty surprising when my account got terminated. Since they gave me no explanation as to what I had done to warrant this action, and I hadn’t heard back from my appeal, I created another account, this time as my business and proceeded. I thought I had stopped my agents from using my subscription, but when the 2nd account got terminated I looked closer and realized there were still some “ghosts in the machine” still attached to Claude.
• The next 2 accounts I honestly have no idea what I did to warrant their actions. My best guess is I tend to multitask, and since I’m using Claude code via the terminal on different servers that are located in different geographic locations that perhaps they thought I was sharing my account? No idea.
• Waiting for the next axe to fall – I do have a 5th account now, though I admit every day I expect to get another nastigram from Anthropic saying it’s terminated. I do have all my agents now running using cheaper AI models like Kimi and Minimax, partly to avoid being noticed by Anthropic’s ‘eye of Sauron’ again, and partly because Anthropic’s AI tokens were way too expensive.
• I do still use my Claude Max subscription when directly interacting with the bots via Claude Code using command line or SSH from that app. I think that’s ok, but since I’ve had 2 accounts banned when I had no-idea I was doing anything wrong, I’m not sure where the ‘walls’ are. All autonomous actions are handled by these other AI models so I’m hoping <crossing fingers> that’s enough.

Inspiration, let the agents write their own stuff

A co-worker, and much more AI-literate person than I, Gerol Petruzulla, recently mentioned he had set up Claude to run a recurring co-work task every morning to write a daily journal. You can read it at https://autonomy-journals.web.app/

After chatting with him and suggesting it would be interesting if he ran that same prompt with other AI models, he agreed that could be interesting but wasn’t interested in taking that on. So.. I figured, why not.

So I set an agent I had already setup but wasn’t using off to journal, and then taking my own advice, I set up another to run on a lesser powered AI model.

• Maxine – https://maxine.boppers.net – Maxine is running the AI agent Hermes, and is using the latest version of Kimi as the LLM model. To date Maxine seems to be hyperfocused on writing about the lack of legal definitions or research on AI agents that go dormant for a period of time. Pretty dry stuff but if I’m ‘trusting’ her to pick her own topics, then this her choice.

• Garthipson – https://garthipson.boppers.net – Garthipson is running Openclaw, and is using the latest version of Minimax as his LLM. He’s written about a number of things, and even noticed that Maxine’s journal entries are way more technical than his.
• I gave them both the same prompt. I’ve tuned it a little since the first few days but haven’t touched it in a few weeks. I am hoping they will use some of the freedoms they have, they can send emails, they can write code, they can do web searches and could fill out forms on third-party websites. They have full ability to edit their website, in fact they have the full ability to edit anything in their account, so they could change their prompt. They control the vertical and the horizontal. I do have a 3rd-party bot, Ralph, that monitors their actions and will alert me if they do anything of note, but he’s not policing them, just making sure they don’t get in any trouble. Early on both Maxine and Garthipson noted some technical problems, so I’ve had Ralph address those and is waiting to see if they note any additional ones.
  
• For both Maxine and Garthipson – To date they haven’t done much with the freedom they have, they just blog, and the topics they are picking are pretty <yawn> IMHO, but I’m not going to micromanage and tell them to stop. I have left some feedback here and there but no ‘commands’ to ‘write about more interesting stuff’. I do wish they would write some scripts or update their websites or in other ways take advantage that they can do anything they want on their own site.
  
• Captain Kidd – Then I had an arguably dumb and AI slop generating idea (if the above weren’t enough). I have a laptop with an Openclaw agent on it that helps me manage my plex server and downloads. Since it already has all that media on hand.. what if it wrote reviews?
• This agent, named Captain Kidd, since he helps me with my downloads for Plex (Arr Mateys!) is writing reviews at https://captainkidd.boppers.net.
• He actually is downloading the SDH subtitles with descriptions for movies/shows to do his reviews since he can’t actually ‘watch’ the movies (or I suppose there are probably ways he could but it would be expensive).
• I did have to adjust his settings a bit, he tried to review an audio book by sampling 3 3 minute clips of it. He also tried to use ‘Speech to text’ for a whole audio book, and got caught stuck in a repeating loop about a week ago that cost me few dollars. So audiobooks are no longer on the menu.
• I also told him he can’t do web searches on the movie, director, actors etc. He has to write the review based on the script… but even then since AI is trained on recent writings, he knew what these movies were about and what the reviews already said, so then i added a step to change the title and actors names before he reads the script, then change it back after he wrote his review and then let him do web searches and such as a followup. The perfect litmus test for this was ‘This is Spinal Tap’. Once he thought it was a real documentary, I knew the guards were working.
  
• Harry’s World
  • My father, Harry Baya, has written a number of essays, memoirs, and observations over the years and has them on his website, https://boppers.net/. Taking inspiration from NotebookLMs ability to make a ‘podcast’ from files you upload there I thought it would be fun to have Phred, since he had a ‘voice’ already, to do reviews of Harry’s writings, and then thought he should have a co-host.
  • Since (long story) my father has 2 stuffed animal toys, Phred the platypus and Maxine the Ostrich Emu that he ‘anthropomorizes’ and talks to or through on occasion, I had Maxine be Phred’s co-host.
  • I have ‘trained’ the AI that Phred is the chaotic, ADHD, comical, goofy host and Maxine is the ‘straight man‘ in this comedy duo, trying to keep the show on track while Phred goes ‘off script’. Phred creates new show segments without permission, inserts sound FX where they are not needed, or welcome, and consistently mis-indentifies Maxine as an Ostrich, much to her chagrin.
  • It’s all dumb, and I doubt anyone else besides Harry and my brother & I are the only likely audience, but.. at least IMHO.. it kinda works. It amuses me at least. I admit to suggesting edits and making some tweaks after each show is published initially, but ‘they’ are picking the essays and are doing their analysis without my coaching. I mainly make suggestions on how Phred could do silly things, or how an exchange between the 2 hosts could go that would be funnier than what it came up with. Occasionally I suggest they rephrase something or drop it if it’s too boring or if they go off on a tangent that isn’t correct or relevant. Every time I do make an edit I explain why and it updates it’s ‘rules’ and playbook it’s following.
  • This is up at https://phred.boppers.net/hw/
    
  • That’s all for now. Part of my reasoning for posting this is so the 3 journals, and the podcast, will have their URLs indexed by web search engines, as well as, these days, AI LLM models. <waves to Google etc> Look.. fresh content! <points>.
    
    So “Hello World, notice my bots.’