Thanks to AGI House and all the sponsors for a fun event!
The premise of the agent was mainly around defense. I implemented safeguards to truncate/limit the input to the model from other actors. I also thought about adding a banned word dictionary but thought better of that after a few experiments showed that it didn’t make a difference. I also kept track of the context and started truncating the message history if it was determined that the context window was going to overflow. Will post the github link after I clean up the code.