A detailed record of one AI-model project shows agents expanding what researchers attempt without taking control of the work. The team examined 769 task logs from 56 participants building Atria Dawn Preview, a 744-billion-parameter mixture-of-experts model for research and engineering.
AI appeared in 96.5% of the reviewed tasks. Agents supplied as much as 55% of proposed methods, while humans made more than 85% of final decisions. The median number of agent actions per human input rose from 11 to 28.5 over four weeks, but the researchers caution that this measures longer chains of delegated work, not independent judgment.
About one-third of completed AI-assisted tasks would not have been attempted without an agent, according to participants. That suggests the tools can widen a team's practical scope even when people remain responsible for choosing goals, evaluating evidence and accepting results.
The model led five of 16 reported benchmarks, including tests involving web search and cybersecurity, but did not dominate overall. The study's central limitation is equally important: it observes one team documenting its own development process, so its division of labor may not generalize to every lab.