Research runs (newest first). Each run is a self-contained snapshot of one execution of the multi-agent pipeline.