The rapid progress of large language models (LLMs) enables them to operate in general‑purpose harness environments, where LLM‑based agents directly manipulate sandbox file systems and shells. Existing post‑training pipelines for tool‑using agents largely rely on an external teacher model: a stronger model creates tasks, gathers trajectories, the target agent is trained once, and evaluation results are consumed by humans, leaving no closed‑loop self‑improvement.
ClawLoop introduces a self‑sustaining framework that eliminates the need for an external teacher and minimizes prior dependencies. It closes the loop with three phases:
- Task Synthesis: generates a diverse pool of tasks via two complementary construction routines, applies diagnosis‑conditioned difficulty growth, and validates tasks both without and with the agent.
- Training: updates the agent using supervised fine‑tuning (SFT) and/or reinforcement learning (RL).
- Diagnosis and Feedback: analyzes rollout behaviors to produce a capability analysis report, which guides the next round of task synthesis.
Using Qwen3.6‑27B as the loop model, ClawLoop consistently improves performance on SkillsBench, Claw‑Eval‑Live, and ClawGym‑Bench. The iteratively refined tasks also transfer to smaller Qwen3.5‑4B and 9B models, yielding gains across all benchmarks and confirming that data quality is independent of the loop model.
Review