Which agent patterns loop most often?

Retry feedback with no bound on it, the same call retried again after a failure with nothing stopping it, is the single most common shape: it showed up in 25% of a study’s 68 confirmed infinite-loop cases across 6,549 open-source agent repositories. Tool-call iteration without a bound is close behind at 23.5%, and multi-agent chat with no cap on turns is third at 20.6%.

Two more patterns cover most of what’s left: a workflow loop with no effective bound at 13.2%, and one agent re-entering a message it already sent at 10.3%. Together the top three patterns account for most of the confirmed cases, ahead of the runner, delegation, or evaluator feedback loops that make up the remainder. That’s a different cut of the same data from why a retry limit doesn’t always stop a loop: that finding is about where the bound is missing, this one is about what the repeating call itself looks like, and the same trace can be read both ways, stuck with nothing ever throwing an error to flag it.