What's the one check you wish you did before running your agent for real?
Instead of chasing better prompts, this thread collects the checks people wish they had done before letting an agent loose: does it stop and ask when unsure, what happens when a tool fails twice, is there a gate before it sends or deletes anything, and does 'done' actually mean done. The useful answers add two more: a hard kill switch on spending, and making the agent show its receipts before you believe it finished.