A study co-authored by researchers at Fudan University analyzed human-AI collaboration during the development of Atria Dawn Preview, a 744-billion parameter mixture-of-experts model. Tracking over 700 task logs across 56 participants, the research documented that AI agents were used in 96.5% of tasks, with agent actions per human input growing from 11 to 28.5 over four weeks. Despite this heavy delegation, human decision-making remained dominant throughout the development lifecycle.

The empirical data shows humans made 85.5% of final decisions regarding methods and parameters and 93.4% of decisions concerning project goals and scope. AI agents primarily served as proposal generators, accounting for up to 55% of suggested options depending on the decision type, but held single-digit percentages in final decision authority. When task issues arose, 76% of interventions required human context clarification or diagnostic assistance.

Notably, participants rated one-third of completed AI-assisted tasks as entirely infeasible without agent support, demonstrating that automation expanded operational scope rather than merely speeding up baseline execution. Researchers concluded that human judgment remains the primary operational bottleneck as AI transitions into a execution partner role.

Why it matters

  • AI agents act as capability expanders for engineering teams, enabling complex tasks that would otherwise remain unattempted.

  • Human oversight remains essential in agentic workflows, with judgment and diagnostic context serving as critical bottlenecks.

  • Founders should design agentic tools around ‘AI proposes, human selects’ workflows to match effective engineering practices.

Source: the-decoder.com