Microsoft Explains Why AI Agents Get Stuck on Long Tasks

Microsoft Research explained why autonomous AI agents get stuck on long tasks. A single agent working alone in one run squeezes itself into a narrow corner of the solution space. The researchers call this an "idea basin," and throwing more compute at the problem does not help it escape. Their method, Fork-and-Flush, splits the agent into parallel branches that share one working folder. Each branch wipes its own chat history and starts reasoning from scratch. Across 13 tasks ranging from a few hours to more than a hundred hours, the average normalized score rose from 0.47 to 0.78. That is a 66% improvement over a single run.

The tests ran on GPT-5.6-terra and GPT-5.5 through GitHub Copilot CLI, not on Claude. Copilot CLI is a command-line tool for running AI agents that write code. The researchers also found that stronger models fall into the same idea-basin pattern. Because of that, they expect periodic context resets to stay useful even as models improve. This gives a tested explanation for why starting from a clean branch can beat one long run at a hard agent task. The paper itself, though, does not call this a universal fix.