AI Agents Context Reset: Fork-and-Flush Lifts Scores 66%

Eight runs of the same research agent on the same task finished with scores anywhere from 68.3 to 100.0, and extra compute did not close the gap. For teams running long-horizon agents, a Microsoft Research study points to an unexpected constraint: the agent’s own chat history, which keeps steering it back to ideas it has already tried.

What Happened

The Decoder reports on a Microsoft Research study of “autoresearch agents”: systems that propose a solution, evaluate it, and refine it from the feedback, over and over. The paper, Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents, found that independent runs plateau at substantially different scores, with gaps that persist even after considerable additional compute.

When the researchers mapped every candidate solution by functional similarity, each run’s solutions almost always stayed in a small region of the map. The authors call these regions “idea basins.” According to The Decoder, agents spent much of their remaining budget refining solutions inside a less productive basin, and the pattern held even when the agents periodically compressed their conversation histories.

The proposed fix has two parts. “Forking” splits the agent into parallel branches that each inherit the workspace and every generated file. “Flushing” wipes each branch’s chat context while keeping that workspace, so it starts a fresh conversation on top of the prior work. After a fixed horizon, the paper’s abstract says, the agent continues from the highest-scoring branch.

Why It Matters

The results came from 13 long-running tasks, including polyomino packing, job-shop scheduling, cryptanalysis, compressing a CIFAR classifier, and optimizing a VLIW kernel, with individual runs lasting 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 Decoder reports that Fork-and-Flush reached a min-max-normalized average score of 0.78, against 0.47 for single runs and 0.54 for best-of-4. That is a relative gain of 66.0% over single runs and 44.4% over best-of-N. It had the highest mean score on 12 of the 13 tasks, and run-to-run spread fell, with normalized standard deviation dropping to 0.16 from 0.26. Clearing the chat context without branching improved results by 22% on its own.

The mechanism is the useful part. A branch that keeps its parent’s chat context tends to return to the same basin, because the earlier reasoning steers its next step. The workspace carries the progress, and the conversation carries the bias.

What This Means for Builders

Most agent harnesses treat the transcript as the agent’s memory and work hard to preserve it, through summaries, compression, and ever-longer context windows. This study suggests that for open-ended search, the transcript can also be a constraint. It is consistent with earlier findings that compressed context drops instructions, and with the push to build autoresearch loops that run for days.

The researchers also argue the effect will not disappear with better models. In tests on MAX-3-SAT problems across Qwen3.5-4B, 9B, and 27B, variance between runs peaked on tasks at the edge of a model’s abilities. Since open research problems sit at that edge by definition, the authors expect regular context resets to stay useful as the edge moves.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.