Agents stop waiting for an external teacher: ClawLoop has them write tasks, diagnose failures, and retrain in the same sandbox

Training for harness-based agents no longer needs a stronger external teacher model, because the agent synthesizes its own tasks, collects its own trajectories, and diagnoses its own weaknesses inside the same sandbox it is trained to operate in. Earlier harness-agent pipelines were one-shot and teacher-dependent: a stronger external model synthesized tasks and collected trajectories, the target agent was trained once, and evaluation results went to human researchers instead of back into the loop. With Qwen3.6-27B as the loop model, ClawLoop reaches 57.2 on SkillsBench against 52.6 for ClawGym and 46.3 for the base model, 76.9 on Claw-Eval-Live against 74.6 and 69.3, and 68.4 on ClawGym-Bench against 61.9 and 67.6.

Each iteration runs three phases: task synthesis, model training, and diagnosis and feedback, where the diagnosis report becomes input to the next round of task synthesis. Task synthesis starts from two complementary routines, one building workspaces from a capability catalog and the other decomposing professional domains that come with tools and a database, then applies difficulty growth and two-stage validation. Training applies self-distilled SFT on successful trajectories and GRPO on tasks that pass validation; each iteration synthesizes 1024 tasks, collects 8 rollouts per task, and keeps only tasks whose pass count falls between 1 and 5.

Tasks produced by the loop transfer to smaller models, indicating that data quality does not depend on the capacity of the loop model. ClawGym relies on external teacher models such as MiniMax-M2.5 and GLM-5.1 for both synthesis and trajectory collection, whereas ClawLoop uses no teacher model at all. Qwen3.5-4B and 9B trained on ClawLoop-generated data improve across all three benchmarks: the 4B model rises from 32.4 to 48.0 on Claw-Eval-Live, and the 9B model from 58.3 to 71.7.

The framework runs inside the sandbox environment provided by OpenClaw, with the same harness used for training rollouts and evaluation to keep the distributions matched; the loop model is Qwen3.6-27B, each iteration synthesizes 1024 tasks, and the loop runs for three iterations. The three external benchmarks are never used for training, task selection, or difficulty scheduling, and the diagnosis phase observes only the held-out synthesized split. It is aimed at readers working on agent post-training and task synthesis.

Performance improves steadily in early iterations but saturates around the third iteration on Claw-Eval-Live and ClawGym-Bench, while SkillsBench continues to benefit from additional iterations. The admission rate does not decay monotonically, falling from 65.6% in the first iteration to 45.8% in the second and recovering to 55.4% in the third, with trajectory length tracing the same shape, which indicates the difficulty-growth routine overshot and then corrected. The 4B and 9B models never run their own diagnosis phase and instead inherit a curriculum shaped by the 27B model's weakness profile, so their gains should be read as evidence about the curriculum rather than as instances of the full loop.