Giving an AI agent more time and computing resources does not guarantee a better solution. Microsoft researchers found that agents working on long research tasks can settle into a narrow set of ideas and keep refining them, while other runs of the same agent find stronger approaches. Their proposed remedy is to restart the conversation without discarding the work already saved.
Why an agent can get stuck
The study examined “Autoresearch agents”: systems that repeatedly propose solutions, evaluate the results, and use feedback to revise their next attempt. In eight independent runs of the same agent on the same task, final scores ranged from 68.3 to 100.0. The differences remained even when the runs received substantially more compute, according to the paper.
To understand why, the researchers mapped solutions from the runs so that similar ones appeared close together. Within an individual run, solutions tended to remain in a small area. The researchers call these areas “idea basins.” An agent can spend much of its remaining compute refining work inside one such basin, even if another run has found a more productive direction.
Compressing the conversation history at intervals did not resolve the pattern. The study’s explanation is that earlier reasoning continues to steer the agent toward similar next steps. In practical terms, extra effort may improve an approach without helping the agent discover a different one.
Fork work, then clear the conversation
The researchers’ method, called “Fork-and-Flush,” has two parts. First, the agent forks into K parallel branches. Each branch receives the existing workspace and generated files, so it can continue from the work already done.
Then each branch is flushed: its chat context is cleared, and it begins a new conversation. The reset removes the prior reasoning that may have narrowed the search, while saved code, files, and results remain available. This gives branches a chance to explore different ideas without requiring them to start the task from scratch.
A branch that retains the parent process’s conversation can be pulled back toward the same idea basin. With a fresh conversation, the agent can use the inherited workspace as a foundation while choosing a different route. The distinction is between preserving the accumulated work and preserving the conversation that shaped the previous search.
Higher scores on the same compute budget
Fork-and-Flush was tested on 13 long-running tasks, including polyomino packing, job-shop scheduling, cryptanalysis, compressing a CIFAR classifier, and optimizing a VLIW kernel. Individual runs lasted from several hours to over 100 hours. The agents used GPT-5.6-terra and GPT-5.5 through the GitHub Copilot CLI.
On the same compute budget, the method achieved a min-max-normalized average score of 0.78. Single runs averaged 0.47, and best-of-4 averaged 0.54. The paper reports relative gains of 66.0 and 44.4 percent, respectively. Fork-and-Flush had the highest mean score on 12 of the 13 tasks.
The results also point to more consistent outcomes: normalized standard deviation was 0.16 with Fork-and-Flush, compared with 0.26 for single runs. Clearing the chat context without branching improved results by 22 percent. In these experiments, a reset contributed value on its own, while branching added a way to explore multiple fresh directions.
Resets may matter at the edge of a model’s abilities
The researchers also tested MAX-3-SAT problems with Qwen3.5-4B, 9B, and 27B. Variation between runs followed an inverted U-shaped curve: it peaked on tasks near the edge of a model’s abilities, roughly halfway between random-guessing performance and the reference solver’s performance.
The authors argue that open research problems, by definition, sit near the limits of what models can do. They expect the variation that Fork-and-Flush uses to persist as models improve, with the boundary moving toward harder problems. If that expectation holds, context resets could remain useful for research agents even as their underlying models become more capable.
The findings suggest a simple design lesson for long-running AI work: the conversation history is not always a neutral record. When it narrows an agent’s search, keeping the workspace and starting a fresh conversation may open paths that continued refinement would miss.