After trying several small local models, I chose this combo for now:
- llama3.2:3b-instruct
- phi4-mini
- qwen2.5:7b-instruct
All running locally via Ollama — no cloud dependency.
The game itself (classic Hangman with two competing agents and a host agent) is actually quite straightforward once the plugin functions are in place.
The real challenge is the prompt engineering (together with the model selection this takes most of the time) Getting the agents to:
- consistently call the right functions in the right order
- actually read and use the tool output instead of hallucinating
- sending back feedback of what they are doing
…turns out to be more complex than writing the game logic.
Current status: agents are starting to behave, they guess letters, sometimes even strategically, but it still needs a lot of hand-holding via very explicit, repetitive instructions.
Hoping to have a short demo clip of them actually playing against each other by next week. I will put the source in my Github repo once it's done.
Still a long way to go, but every time I see them make a smart guess instead of going silent, it feels like real progress.