Atria Dawn's own team rated a third of its finished AI-assisted tasks infeasible without the agent

A task record here is an account of work already done. The paper says the records separate who proposed a method from who selected it, and who diagnosed a difficulty from who carried out the revision ^[9]. The one-third number comes from asking the participants afterwards: they were asked to evaluate completed tasks under comparable conditions, and rated about a third of the AI-assisted ones as infeasible without AI under the same scope and resource constraints ^[7][8]. The people doing the rating are the people who shipped the agent, so the survey leans. I don't think that makes the answer wrong.

The denominator is the set of completed AI-assisted tasks, which the paper does not size ^[15]. So the fraction cannot be turned into a count of tasks, and it cannot be lined up against another team's project without knowing how much of that project ran through an agent at all. Across 56 participants, the 769 records average about 14 each ^[13].

The benchmark side is a claim about 16 harnesses. Atria Dawn Preview reports the highest score on five of them ^[3], which leaves 11 where something else scores higher ^[14], and the authors say performance varies across domains ^[4]. For those five to predict anything about your queue, your work would have to be distributed like the benchmark set and your scaffolding would have to resemble theirs. The process data is the more portable part of this paper, in kind if not yet in strength: it describes a division of labour, and another team can check that against its own logs.

That division is the finding I would try to copy. In the records, agents frequently initiated approaches and executed changes, while humans concentrated on evaluation, selection, and steering, keeping most final decisions ^[10]. The authors call this a move from task-level execution to project-level partnership ^[16]. It was observed with a 744-billion-parameter mixture-of-experts model in the loop ^[5], and it was produced by a training pipeline the paper calls a Verifiable Experience Pipeline, which ties tool-mediated interactions to executable environments and externally verified outcomes ^[2]. I would not expect the same split from a small local model in that seat.

On oversight the authors are explicit. "Progress toward more autonomous AI research must therefore advance both the capacity for discovery and the capacity for meaningful human oversight, preserving accountable human authority over the risks and direction of continued development," they wrote ^[11]. Correspondence for the paper is listed to two Fudan University addresses ^[12].