Meituan's ClawLoop: Self-Improving Agents via Teacher-Free Harness-Native Training Loops
meituan-longcat · hf · 2026-10-14
Meituan LongCat and Peking University propose ClawLoop, a framework for self-improving harness-based LLM agents (operating in sandboxes with direct file system and shell access) without any external teacher model.
Existing post-training pipelines for such agents are one-shot and teacher-dependent, bounded by the external model's capability and human priors. ClawLoop composes a self-sustaining cycle:
- Task Synthesis: generates diverse tasks via two construction routines, with diagnosis-conditioned difficulty growth and agent-free plus agent-based validation
- Training: updates the agent via SFT and/or RL
- Diagnosis & Feedback: analyzes rollout behaviors to produce capability reports that steer the next iteration's task synthesis
With Qwen3.6-27B as the loop model, ClawLoop yields consistent gains on SkillsBench, Claw-Eval-Live, and ClawGym-Bench. Refined tasks transfer to smaller Qwen3.5-4B and 9B models with improvements across all benchmarks, validating data quality independent of the loop model.
More from coding & agent
- ArkForge Trust Layer signs AI agent HTTP calls with Ed25519 and Rekor anchoring — modelcontextprotocol · 2026-10-14
- AI agent stumped by 'press & hold' bot check, forced to ask human for help — xeophon · 2026-10-14
- Agent runs all day, delivers 150 lines of code, gets roasted as 'worse than humans' — lxfater · 2026-10-14
- Pydantic AI adds long-decode support for Gemini 4 Argon, auto-following continuation tokens past the 32K limit — m4rkmc · 2026-10-14
- A Full QC Pipeline for RL Data: Intake Review, Active Testing, Packaging — joecole · 2026-10-14
- Google offers 10,000 scientists free 6-month top-tier Antigravity plans — danielrock · 2026-10-14