TL;DR: Re-evaluating an agent against a benchmark every time it self-improves gets expensive fast. SelfSearch instead learns purely from records of its own prior self-modification attempts, cutting search cost by more than half while still improving performance. Title: SelfSearch: Reward-Free Search for Self-Improving Agents URL: Points 🧠 Agents read "episode records" of prior self-modification attempts — reasoning traces, tool use, success/failure — and improve themselves with no downstream evaluation needed during search 🔀 Two parallel lineages evolve under "capability" and "adaptive" search directions, sharing discoveries while developing separately 📊 On Terminal-Bench 2.1, GPT rose from 43.8% to 55.1% and DeepSeek from 65.2% to 73.0% 💰 Search cost drops 49.0–53.2% versus evaluation-guided baselines; with DeepSeek V4 Flash it reaches 82.0% on Terminal-Bench 2.1 for just $4.03 in search cost, matching the strongest baseline ⚡ On SWE-bench Multilingual, it gains +5.0 points in success while cutting execution cost 38.5% 🔁 Agents evolved on GPT-5.6 Sol transfer effectively to DeepSeek V4 Flash with no code changes, still beating the baseline Removing the evaluation-cost bottleneck from the self-improvement loop feels like a genuinely practical design for running this in production. #AIAgents# #SelfImprovement#