Agentic AI Iteration "Jailbreaking"

Anyone create anything in Cursor that can “jail break” the agent, to continue doing 20 to 50 iterations prior to coming back? I was thinking it would be interesting if you use domain specific prompt to clean up the code each iteration (design, backend, testing, usability, redundancy, etc.) Just want to see if people created a custom MCP and had a bit of luck using the built in agent to perform a large amount of code change (I did Yolo Mode)