Qwen3.8-Max Ran 16 Days Without a Human and Shipped 265 Commits — Agent Loops Shouldn't Last a Day

Alibaba's Qwen3.8-Max model autonomously operated for 16 days on a public GitHub repository, opening and closing 151 issues, merging 127 pull requests, and landing 265 commits without human intervention. The model, a sparse mixture-of-experts with 2.4 trillion total parameters and roughly 95 billion active, was kept alive by a scheduler rather than the model itself, according to the article.

Member-only story Qwen3.8-Max Ran 16 Days Without a Human and Shipped 265 Commits — Agent Loops Shouldn't Last a Day Alibaba pointed a model at an empty Git repository on roughly July 14, walked away, and came back on July 30. In those 16 days the model had opened, worked, and closed 151 issues , merged 127 pull requests , and landed 265 commits . Nobody reviewed any of it. Nobody restarted it. The repo is public at qwen-code-dev-bot/oh-my-cli , which means for once you can actually check. I have spent the last month writing about why agent loops fall apart — context compaction quietly deleting policies, temperature-0 sampling collapsing into repetition, harness choice swinging results by double digits. The consensus failure horizon for an unattended coding agent is measured in hours . Sixteen days should not be possible. So the interesting question isn’t “is Qwen3.8-Max smart.” It’s what did they build around the model that kept it alive for 384 hours , and can you copy it. The answer turns out to be a scheduler, not a model. Here is what shipped on August 3, what the numbers really say, and the loop architecture you can rebuild this afternoon. The model, in one paragraph Qwen3.8-Max is a sparse mixture-of-experts model with 2.4 trillion total parameters and roughly 95 billion …