Most people running agents are running one agent badly. The teams getting real output have made three decisions, and prompting isn't one of them. How much context each agent gets to see. Hand each agent all the context and each produces a mediocre take on the same job. Who gets to call the job done. Let the agent doing the work grade its own work and you get confident garbage fast. What happens on failure. Running a failed run again doesn't teach you anything. Turn a failed run into a…