Fork-and-Flush: resetting an agent's chat context scored 66% higher

Researchers led by Ziyang Cai and Christos Ziakas posted a paper to arXiv on 5 October 2026 arguing that long-running AI agents get stuck because of their own chat history. The Decoder, which reports the authors are at Microsoft Research, summarised it on 12 October. The paper calls the effect "idea basins" and proposes a fix named Fork-and-Flush. All figures below are the authors' own, from tasks they chose, and have not been independently reproduced.

What the paper found

The study looks at what it calls autoresearch agents: systems that repeatedly propose a solution, test it, and use the score to decide the next attempt. Per The Decoder's reading of the paper, eight independent runs of the same agent on the same task plateaued at final scores between 68.3 and 100.0, and the gaps persisted with substantially more compute.

The authors mapped the solutions from every run by functional similarity. Solutions from one run stayed in a small region, which they call an idea basin. They report that agents spend much of their remaining budget polishing a solution inside a weak basin, and that this holds even when the agent periodically compresses its conversation history.

The fix: fork, then flush

Fork-and-Flush has two steps. Forking splits the agent into K parallel branches that each inherit the workspace and every generated file. Flushing wipes each branch's chat context, so it starts a fresh conversation that builds on the saved work but not on the earlier reasoning. After a fixed stretch, the agent carries on from the highest-scoring branch.

The reasoning is that a branch which keeps the parent's chat tends to be steered back to the same basin by its own earlier thinking. The code, files and saved results are kept; only the conversation is lost.

The numbers

The authors tested 13 long-horizon tasks, including job-shop scheduling, cryptanalysis and compressing an image classifier, with single runs lasting from hours to several days. Per The Decoder, the agents used OpenAI models through the GitHub Copilot command-line tool. On an equal compute budget:

• Single run: normalised average score 0.47, standard deviation 0.26.
• Best-of-4 (four independent runs, keep the best): 0.54, standard deviation 0.21.
• Fork-and-Flush: 0.78, standard deviation 0.16, with the highest mean on 12 of the 13 tasks.
• Flushing alone, with no branching: about 22% better than a single run, on 11 of the tasks.

The paper reports relative gains of 66.0% over single runs and 44.4% over best-of-4. The scores are min-max normalised per task, so they measure rank against the other methods, not absolute quality.

Will bigger models fix it?

The authors also ran MAX-3-SAT problems on three Qwen3.5 sizes (4B, 9B and 27B parameters). Variation between runs peaked on tasks near the edge of each model's ability. They argue open problems always sit at that edge, so variation should persist as models improve. That is their inference, not a measured result on frontier models.

What a team can take from it

If you run agents for hours or days on optimisation, tuning or analysis work, do not judge the agent by one run. Run it several times and look at the spread before you set a budget. If you see plateaus, a scheduled reset that keeps files but drops the chat is the cheapest thing to test first.

What the paper does not show: results on everyday business agents such as support or coding assistants, the dollar cost of the extra branches, or how the method behaves when tasks have no clean score to pick a winner.