Recently I saw an interesting article, saying that someone could build a general purpose coding agent in 131 lines of Python code. That's a neat idea, though I'm not sure that this is better than just using Claude Code, especially as the agent still uses the online version of  the Claude model from Anthropic to generate code or perform other tasks. There's a video in the article showing how this code can be used to perform some quick tasks on a computer.

However, the code isn't specific to Anthropic. It can be used with any LLM, and I started doing just that, with a copy of the code from the article, but modified to use a local AI LLM running under Ollama. You can see my repo and feel free to download and play with it. It's expecting a local LLM on 11434.

I'm a big fan of local agents for a variety of reasons, but mostly because I know humans tend to do dumb things. Especially with new technology, and maybe even more especially in development areas.

That includes me.

I'll take shortcuts. I'll give an agent sysadmin on a dev database to try things. I want to be able to experiment, learn, and see what works. I want to learn how to use tools and fail using them. That's how I get better. That's how I get better in sports, in music, and in technology.

And that's not a project I can take time to work on. I don't get to dedicate time to just learn and then go back to work. Work never ends. It's a grinding, constant, continuous treadmill of things I need to deliver to others. I have to learn to experiment around those deliverables when I can find spare moments.

With AI, that means we'll do things that get InfoSec teams to cringe. I get the concerns over data transiting networks and going to who-knows-where to be used who-knows-how-by-others. I appreciate business subscriptions that guarantee that data won't be used, but I also want extra safeguards at times. That means local models. Not necessarily on my laptop, but in my data center.

Plus, that way I (or my org) can control the costs and manage expectations.

I hope local models and local agents catch on, I hope more vendors support them and more organizations are willing to run them. Even in something like AWS Bedrock or Azure Open AI or Vertex AI. Then I can rent the latest and greatest hardware, but have more control over how my organization uses it.