Making an Agentic AI and Humble Pie

Claude Code Opus 4.6 changed my life in 2025. I suddenly was able to generate fully functional scripts and build my own iOS app available on the Apple app store. I created Collective, an alternative social media networking tool. After seeing corporate monopolies ultimately degrade our social structure and human lives, I deleted all of my social media accounts. No regrets. I was convinced my life would be over and I would die alone and never work again if I did it. To the contrary, everything is better with less enshittification. All my super left friends said they wanted a social media alternative... but mine proved unpersuasive. In fact, no one used it. Everyone stayed on Meta, LinkedIn, Spotify, etc which was really weird because the easiest thing to do to align their existence with their proclaimed values was just stop using Instagram... but I didn't understand how complicated it all has become. I realize now how ideals have been infected, compromised, and sink far below the processing power of AI. I don't belive people can democratically solve their way out of this potential train wreck the way things are. Hyper capitalism appears to be catapulting towards a collapse.

Entire Figma design I wired to prototype then exported as a template for a javascript based app w/Fire console

Fast forward, I am going to university in Montreal, Quebec to learn skills to address change from the source. I am deep in machine language and algorithms, while also learning how to fluently code in Python. Not long ago, I suddenly burned out of tokens in one of my Claude Code sessions. How was my usage limit already reached? Within the week and from reddit, I understood what was happening. Enshittification. Artificial intelligence is an extremely important and valuable tool that cannot afford this outcome. I quickly began pulling open source models and got to work on creating my own agent.

I had experience slowly working my way up to this point, where the training wheels really came off. I landed with the Qwen3.6-35B-A3B, Q4 model. Ollama was not happy with it so I got a GGUF version to run with Llamma.cpp. That wasn't great either so I tried LMStudio. That was kind of the sweet spot. I thought that the GUI would slow my model down but its ability to optimize the MoE (mixture of experts) nature of the Qwen 35B model actually resulted in faster speeds. This is because the model will only load the relevant subset of parameters needed at a time. This configuration page is a good reference. Something I did not plan for was the massive KV cache that my model would also have. Instead of operating at 24GB or so of VRAM, it was blowing up to like double that. My computer froze and I rebooted realizing I had to keep the ollama server off.

I'm using the OpenAI api protocol to my local server through LMStudio and it's running really smoothly. By IT, I mean my first little AI Agent I made from scratch. This experience was like dropping into a 5 day k-hole. I got humbled real fast when I had to start getting deep into Pylance and Langgraph by myself. I couldn't just have Claude Code whip me something up, but had to really understand what the hell I was doing. I stepped back and started with a tutorial thanks to Manoj Aggarwal. This is such a simple base though. I wanted to work my way up to something at least remotely in the ballpark of Claude Code (Opus 4.6). So I kept building on it. I'm really proud of myself for getting in there and developing a functioning agent. I used an image from the internets to represent it. I also am working on developing a personality for it since it has persistent memory.

The face of my agent in my mind. It's a ghost in the shell kind of agent, obviously.

Currently it is only CLI accessible and a main python script that imports tools and prompts from others. The core loop (graph) is composed of these nodes : Planning (approval req)-->Research-->Coder-->Executor-->Validator. I used Ollama to stream the AI agent's thinking which was incredibely helpful for debugging. I also realized if you get it good enough in execution to validation, even if it can't get to your solution in that run, it can provide critical feedback for debugging. That was the most rewarding part I think, when my agent ran all the way through, failed, but was like "here is why I think I failed so try to do x." It was incredible.

So now I have my first autonomous agent with sandboxed tools that must be specified in the planning approval phase. It's been working well locally as a research assistant to process large data sets (pandas + numpy + another long day of bashing the keyboard) and then generate scripts for synthesizing them. I love that I understand exactly what is going on under the hood and oversaw every initial line of code to make it. This experience showed me the limitations of AI and importance of humans being in the loop. I less want to have it replace what I do now, and just augment and enhance what I do. I understand the process for it all. If I don't write every line of code, I'm making the flow chart for the algorithm, etc. I'm happy to be both humbled and inspired. In the humility I've learned more and opened my world to a really amazing community of people doing open source work. I hope this is where the change can really start happening.

Working together, we could get off of major platforms, consciously choose reduce invisible unethical AI implementation (Spotify, Amazon, Meta, etc -- caveat I can't get off Google maps or youtube :/). I found a punk group on Mastodon. The open source community can keep us off major corporate APIs and feeding unreasonable and unregulated data centres... perhaps there is hope!