“Full DeLM's speedups were 1.54× with GPT-6-Astra and 2.02× with Claude Opus 5.5. The ablations fell to 0.82–1.09×, which is about single-agent speed.”

Names the cost of agents tripping over each other as "bubbles": waiting on a peer or redoing its work. Evidence that visible claims and status, not agent count, produce the gain. Caveats: selected long-horizon tasks, 10 DeepSWE tasks, and speed is bought with extra compute.