@ferdinand

Ship is here - welcome penguin

Sorry, no ship… but a nice TUI. Do you like it?

Devlog #4 – The Pre-Ship: Penguin - It won‘t fly away!
This will be the last devlog before the launch! Penguin will be released shortly after this devlog, so it’s time for a recap. People often use AI agents for what they call ‘vibe coding’: they give all the work to an agent like Claude Code and think this is the right and smart use of AI. However, I believe that we should approach this technology with more caution and reflection; otherwise, we may begin to delegate highly important tasks without fully comprehending them. This is why I built Penguin — the simplest and, yes, the stupidest AI agent — for those who want to use AI for boring tasks, not fun ones. It’s just one Python script and one JSON file, plus Ollama. That’s it! So try Penguin, and rest assured: it won’t fly away!

Devlog #4 – The Pre-Ship: Penguin - It won‘t fly away!
This will be the last devlog before the launch! Penguin will be released shortly after this devlog, so it’s time for a recap. People often use AI agents for what they call ‘vibe coding’: they give all the work to an agent like Claude Code and think this is the right and smart use of AI. However, I believe that we should approach this technology with more caution and reflection; otherwise, we may begin to delegate highly important tasks without fully comprehending them. This is why I built Penguin — the simplest and, yes, the stupidest AI agent — for those who want to use AI for boring tasks, not fun ones. It’s just one Python script and one JSON file, plus Ollama. That’s it! So try Penguin, and rest assured: it won’t fly away!

Devlog #3 - Just added a memory. Works very well with the qwen3.5-2b model. I started building a tui for the agent… still working on it. I also need to change the system prompt and think about the think/nothink mode.

The last two weeks, I just played around with different models, frameworks, agents, systems, laptops. Then, I had a moment of resignation: small models struggle not with tool use, but with multiple tools. After a few tries, I came back to the absolute basic: bash commands. So I got back to my Agent Friend Framework and just added a tool: execute_bash_command(). That’s it so far. It work’s surprisingly well with the small qwen-model family. Now, I want to try out some extras: finetuning of the model, memory, maybe a very simple web search for troubleshooting, etc. But now penguin is back!

I tested several local models for tool calling performance. Until now, it’s the qwen2.5:3b leading the “board”. To do this, I built a small tool selection(e.g. get_weather, etc.), quite a lot of typing work. But then, I found the agent-friend toolkit in this blogpost(https://dev.to/0coceo/ollama-tool-calling-in-5-lines-of-python-3h5f) and it’s just amazing! I can now get straight into the tool development without hundreds of lines in JSON.