Hi, I just wanted to share some of my research work. I am developing a virtual crew for virtual racing based on AI enabled chatbots that play the role of a Race Engineer, a Strategist and also a Driving Coach. The bots are using a builtin rule engine for deterministic behaviour and some very complex calculations and predictions. But I also use LLMs for conversational duties.
Now the fun part: I use tool definitions in the system part of the prompt to define entry points into the rule system that can be triggered by the LLM.
Also, the rule system can kind of call the LLM in return, thereby creating a constant flow of back and forth between language based reasoning and fact based reasoning. I am convinced, that traditional rule system are the perfect foundation for building agents.