AI News

Large language model agents are increasingly run in harness environments, where they work inside sandboxes and directly use file systems and shells. The authors argue that existing post-training pipelines for these agents are usually one-shot and depend on a stronger external model to create tasks and collect trajectories, while human researchers interpret the evaluations instead of feeding them back into training. ClawLoop addresses this limitation with a self-sustaining, teacher-free cycle. Its Task Synthesis phase creates diverse tasks through two construction routines, increases difficulty based on diagnosed weaknesses, and validates tasks with both agent-free and agent-based checks. The Training phase updates the target agent with supervised fine-tuning, reinforcement learning, or both. A Diagnosis and Feedback phase analyzes rollout behavior and sends a capability report back to the next synthesis iteration. Using Qwen3.6-27B as the loop model, the framework produced consistent gains on SkillsBench, Claw-Eval-Live, and ClawGym-Bench. The refined tasks also transferred to smaller Qwen3.5-4B and 9B models, which improved across all reported benchmarks, supporting the claim that the task data—not only the larger loop model—contributed to the gains.