A 143-author paper says AI agents now write most of a frontier model's own code -- but not the decisions

Shanghai AI Laboratory released Atria Dawn Preview on September 15, a 744-billion-parameter open-weight agentic model under an MIT license, post-trained on top of Zhipu's GLM-5.2 foundation. With it came a 143-author arXiv paper that reads less like a model card and more like a study of the lab's own workflow while building it. The model is trained via what the paper calls a "Verifiable Experience Pipeline," which ties tool-mediated interactions to executable environments and externally verified outcomes rather than static preference data. Across 16 benchmarks spanning research, engineering, and digital work, the paper reports Atria Dawn Preview is competitive with frontier agents and posts the highest reported score on five of them.

What the paper measured

The more unusual part of the release is a labor study bolted onto the model card: the team logged 769 task records from 56 participants alongside the agent's own logs, then had participants evaluate the completed work under comparable conditions. Roughly a third of completed AI-assisted tasks were rated infeasible without AI. The paper's own gloss on the division of labor is direct: agents "frequently propose methods and implement revisions, while humans retain most final decisions." The authors describe this as a shift from task-level execution to project-level partnership, with human effort concentrating on what's worth pursuing rather than on execution itself.

Why the caveat matters

Beyond the capability claim, the paper argues that autonomy and human oversight need to scale together, not trade off against each other, so that accountable human authority over the risks and direction of the work doesn't shrink as the agents' share of it grows. A research lab is publishing evidence of its own agents doing a growing share of the research work, then arguing in the same breath that the human sign-off step can't shrink alongside it. Every lab shipping agentic coding and research tools now faces that tension.