New evals show diminishing returns on multi-step reasoning beyond four turns

A recent arXiv preprint analyzes failure modes in autonomous coding agents across 500 tasks. The data suggests that adding more than four reflection steps rarely improves success rates and often increases hallucination. We should reconsider default loop limits in production frameworks.