Clearing an AI research agent’s chat history can help it find better solutions, according to a Microsoft Research study. The researchers found that long-running agents can settle on one approach even when given substantially more compute.
The team studied “Autoresearch agents,” which repeatedly propose solutions, test them and revise their work using feedback. In eight independent runs of the same agent on the same task, final scores ranged from 68.3 to 100.0. The performance gaps persisted with substantially more compute, the paper says.
Agents stayed near earlier ideas
The researchers mapped solutions from all eight runs so that similar ones appeared close together. Solutions from any single run almost always occupied a small area. The authors call these areas “idea basins.” An agent could spend much of its remaining compute refining an approach in one basin while other runs had found better ones elsewhere. Periodically compressing the chat history did not remove the pattern.
The team’s method, “Fork-and-Flush,” splits an agent into K parallel branches. Each branch inherits the previous workspace, including generated files and saved results, but starts without the earlier chat context. A branch that retains the parent conversation tends to return to the same idea basin, the researchers found. The method keeps the best branch and discards the others.
Fresh contexts improved scores in the tests
The researchers tested the method on 13 long-running tasks, including polyomino packing, job-shop scheduling, cryptanalysis, CIFAR classifier compression and VLIW kernel optimization. Individual runs took from several hours to more than 100 hours. The agents used GPT-5.6-terra and GPT-5.5 through the GitHub Copilot CLI.
On the same compute budget, Fork-and-Flush reached a min-max-normalized average score of 0.78. Single runs scored 0.47, while best-of-4 scored 0.54. The paper reports relative gains of 66.0% over single runs and 44.4% over best-of-4. Best-of-4 improved on single runs by about 14.9%.
Fork-and-Flush had the highest mean score on 12 of the 13 tasks. Its normalized standard deviation was 0.16, compared with 0.26 for single runs and 0.21 for best-of-4. On an 11-task comparison, clearing chat context without branching raised the normalized average from 0.45 to 0.55, an improvement of about 22%.
Read nextNew study shows AI research agents overstate results and lack scientific judgmentModel size did not remove run-to-run variation
The researchers also tested MAX-3-SAT problems with Qwen3.5-4B, 9B and 27B. Across all three sizes, variation between runs peaked on tasks near the edge of each model’s abilities, roughly halfway between random-guessing performance and a reference solver’s performance. Runs were more alike on very easy tasks, where models almost always succeeded, and very hard ones, where they almost always failed.
The results suggest that managing chat context can matter alongside the amount of compute used for research agents. The authors argue that open research problems sit at the limits of models’ abilities. As models improve, they expect that boundary, and the run-to-run variation near it, to shift toward harder problems.
