One run stays inside a small patch of solutions
The Decoder reported on October 12 on a Microsoft Research study of agents that propose a solution, evaluate it, and revise it from feedback. Eight independent runs of the same agent on the same task finished with scores from 68.3 to 100.0. More compute did not close those gaps.
The authors plot solutions so that similar ones sit together. Solutions from a single run almost always stay in a small region. They call that region an idea basin. The agent spends much of the remaining compute polishing inside a less productive region instead of moving to better approaches other runs already found. Periodically compressing the chat history does not remove the pattern.
[1]
Branch, and clear the chat
The method is Fork-and-Flush. Forking splits the agent into parallel branches. Each branch inherits the previous workspace and the files already generated. Flushing means each branch loses its entire chat context but keeps the workspace, then continues in a new conversation.
A branch that keeps the parent conversation tends to return to the same region, because the earlier reasoning steers the next step. A fresh conversation can search elsewhere. The chat is gone. The code, the files, and the saved results stay in the workspace.
[1]On the same budget, the best mean is 12 of 13 tasks
The test has 13 long-running tasks, including polyomino packing, job-shop scheduling, compressing a CIFAR classifier, and optimizing a VLIW kernel. A single run lasted from several hours to more than 100 hours. The agents used GPT-5.6-terra and GPT-5.5 through the GitHub Copilot CLI.
On the same compute budget, Fork-and-Flush reached a min-max-normalized average score of 0.78, against 0.47 for single runs and 0.54 for best-of-4. The relative gains are 66.0 percent and 44.4 percent. It had the highest mean score on 12 of the 13 tasks. Normalized standard deviation fell from 0.26 for single runs to 0.16. Clearing the chat without branching improved results by 22 percent. One task was not the highest, so this is not a win on every task.
[1]