Researchers propose ClawLoop, a self-improving agent training framework for harness-native environments

ClawLoop targets harness-native agents—LLM-based systems that operate in sandbox workspaces and interact directly with file systems and shells—where post-training pipelines are scarce and typically one-shot and teacher-dependent. Existing approaches rely on stronger external models to synthesize tasks and collect trajectories, train the target agent once, and then evaluate results manually, leaving no room for self-improvement.

The proposed framework replaces this dependency with a self-sustaining cycle composed of three phases: Task Synthesis, Training, and Diagnosis and Feedback. Task Synthesis generates a diverse pool of training tasks using two complementary construction routines, applies diagnosis-conditioned difficulty growth, and validates tasks with both agent-free and agent-based checks. Training updates the agent via supervised fine-tuning (SFT) and/or reinforcement learning (RL). Diagnosis and Feedback analyzes rollout behaviors to produce a capability analysis report that feeds back into the next iteration’s task synthesis.

In experiments using Qwen3.6-27B as the loop model, ClawLoop yields consistent improvements on SkillsBench, Claw-Eval-Live, and ClawGym-Bench. The authors further demonstrate that iteratively refined tasks transfer to smaller-scale models—Qwen3.5-4B and Qwen3.5-9B—with gains observed across all benchmarks, indicating that data quality generalizes beyond the loop model.

The paper positions ClawLoop as a step toward scalable, self-improving agent training in harness environments, addressing limitations imposed by external teacher models and human-defined priors.