Ollama-based Multi-Turn MAS Natural Language Conversation Prototype for Scientific Offline Testing

Scientific research that involves human participants testing demands systematic and controlled study-specific procedures, compliant with local data and research ethics regulations. As interest in human- to-AI-agents social dynamics grows, there is a need for new systems that are both easily controlled, flexible, and ethically compliant. Open-weight large language models (LLM) allow for new types of conversational agents to be used by the scientific community. However, most of these platforms are hosted on online servers and offer limited control over experimental conditions and participants’ data privacy. We have developed a prototype for an experimental testing tool that could ease agent-to-agent and human-to-AI-agent communication when testing complex hierarchical multi-agent interactions. Our user-friendly toolkit provides a platform and a local Rust application using Ollama, allowing integration of any available offline LLMs, such as Gemma3 and GLM-4.7-Flash. This demonstration of agent-to-agent communication showcases the toolkit's resilience under different LLM models, encouraging its usage by the wider community, as our human-to-agent platform is being developed and tested. We produce an environment where LLM-based multi-agent conversations with several trained models can be tested, and discuss its further embedding in human-to-AI agent research environments that: 1) depend on adjacent computing systems, e.g. OxySoft and Lab Streaming Layer in the same pipeline, and 2) are tailored for ethical and rigorous compliance with scientific data gathering, networking and processing.