Harness-R1: Learning to Edit Executable Runtime Harnesses from Agent Failure Trajectories

Harness-R1 treats an AI agent’s runtime harness—not just its model weights—as a trainable object. A dedicated 9B harness engineer receives batches of failure trajectories from a frozen Qwen3.5-9B agent, then generates validated executable patches across four lifecycle hooks: initialization, pre-decision guidance, pre-action mediation, and post-feedback recovery. The engineer is initialized with supervised fine-tuning from GPT-5.5-generated patches and optimized with online group-relative policy optimization, sampling 8 candidate patches per failure packet and rewarding only the target agent’s same-batch task improvement. Across WebShop, ALFWorld, and DBBench, average success rises from 44.3% to 53.6%, a gain of 9.3 percentage points, outperforming fixed or prompted editors including OpenAI’s GPT-5.5, DeepSeek-V4-Pro, Google DeepMind’s Gemini-3.5-Flash, and GLM-5.2. After direct fine-tuning of the target agent, Harness-R1 improves performance further from 59.2% to 64.2%, adding 5.0 percentage points. The learned editor transfers to 20 unseen target models with an average gain of 7.06 percentage points, and sparse evidence generalizes to 1,270 held-out tasks with an 8.9 ± 1.5 percentage-point improvement. The results position harness editing as a complementary learning axis that can co-evolve with model fine-tuning, while highlighting the importance of outcome-verified, failure-conditioned executable interventions over plausible but untested code changes.