Researchers find complex AI-agent setups offer no edge on ML-engineering benchmarks

Holding the model and time budget constant, the authors point to the underlying AI—not extra orchestration—as the main performance driver in the tested setting.