Microsoft boosts AI research agents by resetting context

Microsoft Research has uncovered a major bottleneck in autonomous AI research systems: long-running agents plateau because accumulated conversation context traps them in stagnant reasoning patterns. In tests evaluating independent runs of the same agent on identical tasks, final performance scores varied drastically from 68.3 to 100.0. Rather than exploring novel approaches, agents repeatedly spent compute refining solutions within localized idea basins, even when periodically compressing their conversation histories.

To break this loop, researchers developed a framework called Fork-and-Flush. Under this method, an agent splits into K parallel branches, inheriting existing workspace files, code, and saved results, while completely discarding its conversation context. Tested across 13 complex long-running workloads—including job-shop scheduling, polyomino packing, cryptanalysis, optimizing a VLIW kernel, and compressing a CIFAR classifier—agents executed runs lasting up to 100 hours or more using GPT-5.6-terra and GPT-5.5 through the GitHub Copilot CLI.

With equal compute budgets, Fork-and-Flush achieved a min-max-normalized average score of 0.78, outperforming standard single runs at 0.47 and best-of-4 approaches at 0.54, marking relative score gains of 66.0 percent and 44.4 percent respectively. The technique secured the highest mean score on 12 of the 13 tasks while lowering the normalized standard deviation from 0.26 down to 0.16. Simply resetting chat history without branching boosted scores by 22 percent. When evaluating MAX-3-SAT problems across Qwen3.5-4B, 9B, and 27B models, performance variance followed an inverted U-shaped curve that peaked at the limit of each model's capability.

For AI practitioners and engineers building autonomous agents, these findings demonstrate that context window management requires active clearing rather than endless history accumulation or simple summary compression. Preserving practical assets like workspaces while regularly flushing conversational memory prevents agents from getting trapped in narrow solution spaces, unlocking significantly higher efficiency on hard research problems.