Mᴀʀᴋ VᴀɴᴅᴇWᴇᴛᴛᴇʀɪɴɢ

@brainwagon
428 Followers
198 Following
3.8K Posts
He/Him. Formerly Senior Technical Director, Pixar Animation Studios. K6HX. Telescope Maker. Tinkerer. Woodworker. Programmer. Open source afficionado. Enthusiast for enthusism. Make stuff, and share your knowledge and expertise!
I updated this for 2026.
The Gemma 4 E4B model has definite limitations, but it's kind of shocking how well this works, running the model entirely locally (and using my own agent harness).
I just don't understand the appeal of Rust as a programming language.
Allright, tomorrow I am going to _not_ run Claude code at all, and instead do an old fashioned walk through of all the code in my agent 99 repository. It's now got a bunch of functionality, and I think it would be worth reviewing it personally to make sure I understand it, and see what might be improved.
I updated my agent "99" to accept image attachments, which can be described by the gemma 4 models, which include image understanding.

Okay, I'm pretty much done for the day with respect to tinkering with my aforementioned minimal (ish) harness for AI tinkering. I've placed it in a github repository here: https://github.com/brainwagon/99

I've called it 99, after the character played by Barbara Feldon, Agent 99, in the old Get Smart TV show. It's public domain, having mostly been written by Claude with some guidance by me, but more to instruct me on how these things work. Enjoy.

GitHub - brainwagon/99: A near minimal AI harness, written by Claude Code to help me understand how harnesses like Claude Code work.

A near minimal AI harness, written by Claude Code to help me understand how harnesses like Claude Code work. - brainwagon/99

GitHub
Today's AI exploration: writing my own very minimal coding harness. It's actually not that difficult, just a couple hundred lines of Python to implement the basic loop, tool calls, a SOUL.md file, subagents... My suspicion is that a large amount of code in many agents is trying to manage security. Even then, it seems pretty interesting, if not ready for mass deployment/adoption.

"ollama-python 0.1.2 requires httpx<0.27.0,>=0.26.0, but you have httpx 0.28.1 which is incompatible."

Sigh.

Note to self: time to learn Kicad.

Not gonna lie, I burst out laughing reading the last bit.

Credit to KDE and the source of the photo: https://kde.org/anniversaries/30/