The Pulse
Fresh-Context Agent Branches Lift Scores Across 13 Tasks
A study evaluates “fork-and-flush,” a method that gives parallel agent branches a fresh chat context while preserving their accumulated workspace. Across 13 research and engineering tasks, the authors report higher normalized average scores
AI.info Team · Written with AI ·
“Across 13 long-horizon research and engineering tasks, with individual agent runs lasting up to several days, fork-and-flush outperformed the single-run and best-of-N baselines by a relative improvement of 66.0% and 44.4%, respectively, on the min-max normalized average score under an equal compute budget.”
Ziyang Cai and co-authors, the paper’s authors, write. “Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents”
Keep the workspace, clear the conversation
A research team proposes a way to help long-running AI agents break out of unproductive approaches: preserve the files and accumulated work, but restart the agent with a fresh chat context. The method, called “fork-and-flush,” periodically launches several parallel continuations from the same workspace, then keeps the branch with the best score.
The study appeared on arXiv on October 5 under the title “Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents.” Its authors describe a recurring problem in open-ended work: separate runs of the same agent can settle on different approaches and plateau at different scores, even after spending substantial additional compute. They call the clusters of similar approaches “idea basins.”
Fresh context is only one part of the intervention. Fork-and-flush also explores multiple continuations and selects among them using each task’s score. The experiment therefore tests a combined strategy—not a simple chat reset in isolation.
Four branches, two scheduled resets
For the main evaluation, the researchers used 13 research and engineering tasks, including packing shapes, scheduling jobs, cryptanalysis, optimizing a kernel for a simulated machine and training a compact CIFAR image classifier. Individual runs took from several hours to multiple days.
The agents began with four candidates, then branched again at 33% and 66% of each task’s compute budget. In each later round, four branches inherited copies of the leading agent’s workspace but started with empty chat contexts. The branches explored for a set budget; the system promoted the highest-scoring one and discarded the others.
The comparison used matched compute budgets. The single-run baseline spent the full budget on one line of work; best-of-four divided it among four independent runs and kept the top result. The team used GitHub Copilot CLI version 1.0.73 as the coding agent, with GPT-5.6-terra for 11 tasks and GPT-5.5 for the kernel and CIFAR tasks.
The reported gain is an average, not a universal win
On the paper’s min-max normalized average score, fork-and-flush reached 0.78, compared with 0.47 for a single run and 0.54 for best-of-four. The authors express those differences as relative gains of 66.0% and 44.4%. Fork-and-flush recorded the highest mean score on 12 of the 13 tasks, but it did not lead every task: the single-run baseline had the highest mean on the cryptanalysis task.
The authors also tested forking and flushing separately across 11 tasks. Flushing alone scored 0.55 on the normalized average, while forking without clearing chat context scored 0.57; the combined method scored 0.71 in that separate comparison. Those results suggest both mechanisms contributed to the outcome, though the ablation used a smaller task set than the main evaluation.
The aggregate score is a normalized average across tasks with different scoring systems, not a percentage improvement in real-world productivity or a direct measure of how much better every individual answer became. The study’s results apply to its selected tasks, model configurations and evaluation harness. The authors say the expense and time required to run long-horizon evaluations limited how many tasks they could include, and that their branches operated independently, without coordination or specialized roles.
A practical test of agent context
The paper’s central distinction is between a project’s accumulated artifacts and the conversation that produced them. The agents inherit the former but not the latter, giving each branch a chance to interpret the same work without following the previous chat’s chain of reasoning. That design makes context management an experimental variable alongside parallel search and score-based selection.
For practitioners, the reported method is a scheduling strategy: periodically create fresh-context branches from a shared working state, give them bounded exploration time, and retain the strongest result. The study’s clearest boundary is equally concrete: it evaluates 13 tasks, and its authors identify broader task coverage and coordinated branches as future work.