Create a proof of concept for an LLM based agent

[kevinforalignment]

• [ ] Design environment
• [ ] Select open model for tool useage
• [ ] Set up LLM inference model
• [ ] Set up tool usage to interact with environment
• [ ] Set up agent environment

[kevinforalignment]

Design environment:

• Model: open agent model trained on tool usage
• Actions: text response which is parsed to call tools, as defined by a game master
• State: rolling conversation summary (summary of previous summary + rolling conversation history) + rolling conversation history (last x tokens)
• Environment engine: parses test response and calculates rewards

[kevinforalignment]

Make it into a game that people can play?

[kevinforalignment]

Design environment:

• Model: open agent model trained on tool usage
• Actions: move, boost harvest, take harvest, share harvest
• State: observed agent history, claims made about self and others by agents
• Environment engine: parses test response and calculates rewards

[kevinforalignment]

Start proof of concept by using phi-3-mini model and scale up to a bigger model or swap to a different model class as needed. phi-3 is attractive because of it's small size, however, it may not be appropriate due to it's lack of pretraining on function calls. The initial goal is to set up a pipeline quickly and cheaply to experiment with and make adjustments from there.

[kevinforalignment]

Start a new repo for this issue: https://github.com/buildingforalignment/agent-experiments