Dim Neocleous

I have been using Claude code and cursor for the past almost 2 years. Everybody is talking about loop engineering and graphs. I have such a setup with over 10 agents. It used to work pretty well. Over the past month, the more autonomy I give to the agents the more complicated they make the solution thus getting trapped in a loop for hours. They tend to find bugs and add nice to have features that are not even nice to have when the effort is huge and the value is low. Am I the only one experiencing this strange behavior?

For those building or deploying agents, what checks do you use before giving them responsibility for real business work? My concern is that completing a task successfully doesn t necessarily mean the agent respected the business s rules while doing it. Do your evals cover situations where an agent completes the task but also takes an unauthorized action or presents invented information as fact? A concrete example of a test that caught this would be useful. And if something slips through, what happens next? Who gets alerted, who can stop the agent, and how do you investigate or reverse its actions? I d like to hear what has worked in practice, including failures that changed how you test or deploy agents.