ClawLoop: Research Framework Lets AI Agents Self-Improve Without an External Teacher

It is a research framework proposed in an arXiv paper that lets AI agents operating in sandboxed harness environments improve their capabilities through a self-sustaining loop of task synthesis, training, and diagnosis, without relying on an external teacher model.

Traditional pipelines are largely one-shot and teacher-dependent: a stronger external model synthesizes tasks and collects trajectories, then researchers consume the evaluation. ClawLoop feeds diagnosis results back into task synthesis, making the next iteration harder and turning the pipeline into a continuous loop.

Qwen3.6-27B was used as the loop model, with evaluation on SkillsBench, Claw-Eval-Live, and ClawGym-Bench. The refined tasks were then transferred to Qwen3.5-4B and 9B models, which showed gains across all benchmarks.

The work is research-stage with no released production tooling or licensing details, so practical adoption would require reimplementing the framework internally and validating results on the organization's own data first.