Stop treating agents as miracle workers
Users expect agents to work magic without proper context or prompts, then blame the tool when it guesses. The practical fix shared in the thread is to treat a new agent like a new hire: give it context, examples and boundaries, and check its first ten outputs before trusting it.