ClawLoop enables agent self-improvement without external teacher models

ClawLoop, a harness-based agent training method detailed on arXiv, lets file-system/shell agents self-improve in a closed loop without a stronger teacher model by synthesizing and validating their own sandbox tasks, training with SFT/RL, then feeding rollout diagnosis back into the next task-generation round. The iteratively refined task data transfers to smaller 4B and 9B models with benchmark gains, allowing a larger loop model to bootstrap a reusable training-data pipeline for cheaper agents instead of continuous distillation from a frontier model.

arXiv https://arxiv.org/abs/2610.14118 ClawLoop enables agent self-improvement without external teacher models Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. ClawLoop trains harness-based agents in a closed loop without a stronger teacher model: it synthesizes sandbox tasks, validates them, trains with SFT/RL, then feeds rollout diagnosis back into the next task-generation round. For production agent teams, the important shift is that eval failures can become automated training signal instead of offline researcher analysis, making it a plausible pattern for continuously improving file-system/shell agents while reducing dependence on proprietary frontier-model data generation. A harness-based agent file system/shell sandbox can now self-improve without a stronger teacher model, using a closed loop where the agent synthesizes and validates its own tasks, trains via SFT/RL, then diagnoses its own rollouts to steer the next round of task generation—breaking the usual ceiling imposed by teacher-model capability. The practical payoff: the iteratively refined task data transfers to smaller 4B/9B models with gains across benchmarks, so you can use a larger loop model to bootstrap a reusable training-data pipeline for cheaper agents rather than paying for continuous distillation from a frontier model.