More steps, the same kind of measure
The team developing Atria Dawn Preview examined more than 700 task logs from 56 participants alongside their agents’ logs. Across four weeks, the median number of agent actions per human input rose from 11 to 28.5. That ratio counts how much activity follows a human instruction. By itself it says nothing about who chose the research goal or approved a proposed method. The useful move in this study is to count decisions separately from the steps taken to execute them.[1]
Long chains of agent work make the distinction consequential. One human decision can launch more tool calls and intermediate outputs, so the ratio may rise while formal authority stays put. There is another possible reading, too: participants may have retained the final choice while reviewing the options only lightly. The logs identify who made a recorded decision; the action ratio cannot tell us how closely that person examined the underlying work. Atria therefore gives a narrower result than a claim that agents independently directed the research.[1]
The limits of human choices
For methods and parameters, the most common pattern was an agent proposing and a person selecting, at 55.4 percent. People made 85.5 percent of final method and parameter decisions and 93.4 percent of final choices about goals and scope. Participants also judged 151 of 455 completed AI-assisted tasks infeasible without AI at the same scope and quality. That last figure is their counterfactual judgment about their own work, rather than a controlled productivity comparison.[1]
The next useful test would inspect the choice itself. Would a participant choose differently when shown the rationale for each option, test results and contrary evidence? Would another team select the same agent proposals on comparable tasks? Logs from one team make its division of labour unusually visible. Independent teams and prespecified decision criteria would be needed to learn whether the choices are well informed and whether this pattern travels beyond Atria’s own development project.[1]