Microsoft Research paper proposes "Fork-and-Flush" to reset chat context in long-running AI research agents, reporting 66% higher normalized score

Microsoft Research (Cai, Ziakas et al., 2026) published a paper on long-running "autoresearch" agents. It found that repeated runs of the same agent plateau at very different levels (scores from 68.3 to 100.0), which the authors blame on accumulated chat context trapping agents in "idea basins". Their fix, Fork-and-Flush, splits a run into K parallel branches that keep the workspace but have their chat history cleared. On 13 tasks, using GPT-5.6-terra and GPT-5.5 through GitHub Copilot CLI, it scored a normalised average of 0.78 at equal compute, against 0.47 for single runs and 0.54 for best-of-4. Flushing alone gave about +22% on 11 tasks. We have this through The Decoder, with no link to the paper.

Anyone building long-running coding or research agents can try this cheaply: keep the files and results, drop the conversation, and run several branches. The best-of-4 comparison is the useful one, since it suggests the gain comes from the reset and not simply from running more attempts. Impact is still moderate. It is a lab result on a chosen task set, and nobody has shown it holding up in production or on real open research problems.

The claim that resets will stay useful for frontier research agents is an extrapolation from variance patterns, not something the tasks demonstrated. The headline "66% higher" comes from min-max normalised scores on 13 chosen tasks, which can exaggerate gaps, and the idea of clearing context or branching runs is not new. Without the paper, we cannot check how the tasks were chosen, why flushing alone was measured on only 11 of them, or whether the "idea basins" explanation was tested.

1. 01Publication of the paper itself or its code: check task selection, the normalisation method, error bars and the compute accounting behind 'equal compute'.
2. 02Independent replication on other models and agent frameworks, including open-weight ones such as Qwen3.5, and on tasks not chosen by the authors.
3. 03Whether GitHub Copilot CLI or other agent tools add a context-flush or fork feature in the next few months, which would show real adoption.