ClawLoop: Self-Improving Agents via Harness-Native Training Loop
ClawLoop trains harness-based agents by feeding their own rollout diagnoses back into the next task-generation cycle.
The paper frames current harness-agent post-training as mostly one-shot and dependent on stronger teacher models. ClawLoop instead cycles through task synthesis, SFT and/or RL training, then diagnosis and feedback for the next round. With Qwen3.6-27B as the loop model, the authors report consistent gains on SkillsBench, Claw-Eval-Live, and ClawGym-Bench. They also say the refined tasks transfer to smaller Qwen3.5-4B and 9B models, suggesting the generated data is useful beyond the loop model. HF Daily Papers' note
The paper frames current harness-agent post-training as mostly one-shot and dependent on stronger teacher models. ClawLoop instead cycles through task synthesis, SFT and/or RL training, then diagnosis and feedback for the next round. With Qwen3.6-27B as the loop model, the authors report consistent gains on SkillsBench, Claw-Eval-Live, and ClawGym-Bench. They also say the refined tasks transfer to smaller Qwen3.5-4B and 9B models, suggesting the generated data is useful beyond the loop model. HF Daily Papers' note
score 5