Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents
Short summary
This paper identifies that long-running AI research agents get stuck in "idea basins"—localized regions of solution space—causing independent runs to plateau at different performance levels. The authors propose fork-and-flush, a method that periodically branches the agent into parallel trajectories with fresh chat contexts while preserving accumulated work, then continues from the best performer, achieving 66% relative improvement over single-run baselines across 13 complex research tasks.
- Independent runs of the same agent on identical tasks plateau at substantially different scores, with performance gaps persisting even after significant additional compute, suggesting agents explore only localized regions of the solution space rather than broadly searching.
- Fork-and-flush works by forking the agent into K parallel trajectories at episode boundaries, clearing the chat history while keeping the workspace intact, running each for a fixed horizon, then continuing from the highest-scoring branch—this simple intervention requires no architectural changes.
- Across 13 long-horizon tasks (individual runs up to 107 hours), fork-and-flush achieved 66% relative improvement over single-run and 44.4% over best-of-N baselines under equal compute budgets, with artifact embeddings confirming broader exploration through solution space.
Full summary
Researchers at Microsoft found that when you run the same AI agent on a hard research problem multiple times, it consistently gets stuck. Different runs plateau at wildly different performance levels—one might reach a score of 100 while another maxes out at 68—and these gaps don't close even with way more compute thrown at the problem. When they mapped out what solutions each agent found by looking at their functional similarity, they discovered the runs weren't exploring broadly. Instead, each agent was trapped in what they call an "idea basin"—a localized region of the solution space where it kept refining variations on early choices rather than discovering the better solutions found by other runs.
To break agents out of these basins, they tested a simple intervention called fork-and-flush. Periodically, the method splits the agent into K parallel branches. Each branch gets a copy of the accumulated work (the files and code it's built up) but starts fresh with an empty chat history—no memory of previous reasoning. After running all branches for a fixed time, the agent continues from whichever branch scored highest. The key insight is that dumping the chat context while keeping the workspace lets the agent explore different directions without being anchored to earlier decisions.
Across 13 challenging tasks—things like designing compact neural networks, scheduling problems, cryptanalysis, and scientific inference—fork-and-flush crushed the baselines. Individual runs lasted anywhere from hours to several days. Compared to a single run, fork-and-flush delivered a 66% relative improvement in average score under equal compute budgets. Against a "best of N" baseline (just running the agent multiple times independently), it still won by 44.4%. The embedding visualizations confirmed the mechanism: while a standard run stays confined to one basin, fork-and-flush's fresh-context branches actually jump to new regions and continue exploring from there.