Microsoft researchers discovered that long-running AI agents can plateau despite additional compute, with performance gaps persisting even with more resources. The study, focusing on 'Autoresearch agents,' found that agents often get stuck in 'idea basins,' limiting their ability to explore better solutions.
The team tested a method called 'Fork-and-Flush,' which combines branching and context clearing. This approach allows agents to start fresh conversations while retaining prior work, leading to improved results. On the same compute budget, Fork-and-Flush achieved a min-max-normalized average score of 0.78, compared to 0.47 for single runs.
The method was tested on 13 tasks, including polyomino packing and job-shop scheduling, using GPT-5.6-terra and GPT-5.5 models. The results showed Fork-and-Flush achieved the highest mean score on 12 of the 13 tasks, with more consistent outcomes and lower standard deviation.
"Clearing the chat context without branching improved results by 22 percent," said Matthias Bastian, a researcher at Microsoft. This suggests that context resets can enhance the diversity of solutions explored by AI agents.
The study also noted that stronger models may still benefit from context resets, as variance in performance peaks at tasks near the edge of a model's capabilities. The authors expect the variation Fork-and-Flush exploits to persist as models improve.
Microsoft did not comment on the potential for future enhancements or how the method might be applied to other AI systems. The results suggest that regular context resets could remain useful even for advanced research agents.
Source: thedecoder
