TL;DR A method that rewrites successes found under several harnesses into training data a single general harness can reproduce. Training on the raw traces hurts; rewriting them lifts pass@3 from 57.0% to 74.2%. Title: Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite URL: Points 🧩 Different harnesses solve different tasks; the union of three solves 34.3% more than the best single one 📝 A planner turns each success into a runbook, and an executor re-solves the task from scratch under the general harness 🔍 A critic screens every runbook for leaked answers and sends rejects back with a reason 🧹 The model then flags values that can't be derived from the task or environment and discards those traces 📈 Training data grows from 2,001 to 11,094 trajectories, and Terminal-Bench Hard rises from 39.0% to 63.0% ⚠️ Direct SFT actually degrades TB2, with dead-end loops where the model repeats the same action pattern forever 🎛 Even with no training, swapping the harness doubles TBH from 33.0% to 66.0% If you're thinking about feeding your own agent logs back into training, this makes a solid case for a rewrite-and-audit step first. #AIAgents# #SelfImprovement#