Hey all, I wanted to test some ideas for Agentic memory and compaction and context. The project evolved into an Agentic Collective . Now trying to let. them have emergent behaviour with access to a limited set of scientific skills and search tool. Runs on the spark pretty nicely. stil finding bugs and optimizations, feel free to contribute code if you find this an interesting experiment,
Why did you choose to run four different models and only one agent instance per model as opposed to one bigger and smarter model and four (or many more) different agent instances with their unique context/system prompt bias? Would probably yield better results overall.
To be honest, I already had ollama running, so I just picked 4 random models to start with and started to move towards the memoryengine testing. I did notice especially at inception, the context/system and prompt bias was very limited.. still is. there is some divergence as they find different sources of information and internalize it. but I like you idea. will explore it.
I asked Claude.
”So the real design question is: **is your collective’s value in consensus or in productive disagreement?”
- Genuine cognitive diversity, not simulated diversity**
When you run 4 instances of the same model with different system prompts, you’re creating bias variations on a single reasoning style. The underlying weights — how the model decomposes problems, what associations it makes, what it finds salient — are identical. The “different perspectives” are largely theatrical.
Running Qwen, GLM, LLaMA, and Deepseek means the actual inference pathways are structurally different. Deepseek has different training distribution, different RLHF tuning, different architectural choices than Qwen. When they disagree, it’s more meaningful.
Updated: with divergent identities. agentcollective/config.yaml at main · ryanprice/agentcollective · GitHub