Fork-and-Flush branches agent work, retains promising workspaces, and clears conversation histories. Across thirteen long-horizon research tasks compared under the same computing budget, the paper reports a normalized average score of 0.78, versus 0.47 for a single run and 0.54 for choosing the best of four runs. The central idea is to reuse code and experimental results without inheriting mistaken assumptions accumulated in context. This difference should not be extrapolated into a claim of 66% higher productivity in general software development.