Dream-RSI: DeepMind Agents Improve Without Retraining

Google DeepMind has published Dream-RSI, a method that helps AI agents improve their strategy without retraining the model. The agent records its attempt history and then replays those attempts with different settings. The goal is to determine what would have worked better, without launching new costly runs. On a code optimization task, the number of attempts dropped from 550 to 317. Execution time fell from 3,587 ms to 2,931 ms.

For agent pipelines, Dream-RSI is useful in cases where the agent makes many attempts in a row. The baseline method requires 51,200 runs; Dream-RSI finds the strategy in 317. The difference in token consumption is obvious. I would watch whether something like this appears in Claude Code. Agents that optimize code in a loop lose money on exactly this.