DeepMind: Dreaming DeepMind Agents Get Better

DeepMind has introduced Dream-RSI, which allows an agent to remember past attempts and simulate them under various conditions without running expensive new ones. In this way, the agent can decide what could have been done better without running any new (and expensive) attempts. For example, when optimizing code, the number of attempts went down from 550 to 317. Execution time went from 3,587 ms to 2,931 ms.

Dream-RSI is especially relevant for agent pipelines, where an agent performs a large number of sequential attempts. In this case, it takes 51,200 runs to find the optimal strategy via the basic approach, while Dream-RSI manages to do it with just 317 runs. It’s easy to see how much more tokens are spent in this scenario. I’m looking forward to seeing something similar in Claude Code. Losing money on iterative code optimization is standard for such agents.