Microsoft's Fork-and-Flush clears agent chat history to escape AI research plateaus
Overview
Microsoft Research proposes Fork-and-Flush, a method that runs parallel branches of long-running AI research agents with fresh chat contexts while keeping the shared workspace.
The study says that agents otherwise plateau in a narrow "idea basin" despite more compute. On 13 tasks, the method reached a 0.78 normalized average score, compared with 0.47 for single runs.
The authors say regular context resets could remain useful even for frontier research agents. The report covers a single study, and the results come from the authors' own benchmark.
Written by AI from the articles below · updated Oct 12, 1:00 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- The DecoderNewsMicrosoft researchers find clearing chat history helps AI agents find better solutions
AIMicrosoft Research reports that long-running AI research agents can plateau in a narrow "idea basin" despite more compute. The proposed Fork-and-Flush method runs parallel branches with fresh chat contexts while keeping the workspace, reaching a 0.78 normalized average score versus 0.47 for single runs on 13 tasks. The authors say regular context resets could remain useful even for frontier research agents.
Heat trend
Not enough continuous observations to show a trend yet.