Google teaches Agents to 'dream': reviewing past trial and error, so they take fewer detours in the next round.

Beating AI News Flash: Google's research team has released Dream-RSI, which allows an AI Agent, after completing a round of tasks, to use past successes and failures to "dream." The system saves each attempt and its result, then uses these old records to simulate different approaches, such as which path to try first, when to give up, and whether to try multiple options simultaneously. Because the results have already been run before, these simulations do not require re-executing the tasks.

The system selects the better-performing approaches and directly applies them to the next round. The new round then leaves behind more successes and failures, which are again used to "dream" and continue refining the strategy. Round after round, this is what the paper calls "recursive self-improvement." Currently, the underlying model itself does not change; what is actually improved is the Agent's method for "what to do next."

The paper tested this method on 8 tasks, including algorithms, mathematical optimization, and GPU kernels. In one algorithm task using Gemini 3.1 Pro, compared with a fixed approach, Agent calls dropped from 550 to 317, while the final program found also ran faster.