I have been experimenting with multi-agent setups lately, but the biggest challenge is definitely managing tool-call storms when the orchestration logic gets messy. It makes debugging production logs a total nightmare red teaming AI agents for security