- A team including Fudan University researchers studied its own build of the Atria Dawn Preview model: humans made the final call on goals in 93.4% of cases, and on methods in 85.5%.
- Agent actions per human input rose from 11 to 28.5 over four weeks, and the authors warn that longer chains of agent work could leave reviewers only rubber-stamping.
- This is one project, not an industry benchmark. Leaders can ask which approval gates are real decisions, and who signed off on running agents in autonomous modes.
Agent steps per human input rose; the authors do not read that as autonomy
Who decides when AI agents help build an AI model? A team that includes researchers from China's Fudan University put that question to its own project, and The Decoder reported the results on September 27, 2026. The subject was Atria Dawn Preview, an agentic language model with 744 billion parameters. The evidence came from two sets of records: over 700 task logs contributed by 56 people, and the activity logs of the agents those people used. This is one team examining itself, not an industry survey. AI featured in almost every task reviewed.
Over four weeks, the median number of agent actions for each human input climbed from 11 to 28.5. The authors warn against treating this as a sign that agents were becoming more independent. Their explanation is that a single human decision set off a longer run of agent steps, not that agents were taking more decisions themselves.
Work that might never have been attempted
The team also asked contributors a counterfactual: could they have delivered their part of a task, at the same scope and quality, with no AI? Among 455 completed AI-assisted tasks, 151 got a "no." Those 151 were distributed across 27 of the 56 participants, so a small circle of intensive users does not explain them. In these cases the study reads AI as the reason the work existed at all, not as a way to finish planned work sooner.
Goals and methods stayed with people
On what to build and how far to take it, a person made the last call in 93.4 percent of cases. On methods and parameters, humans made 85.5 percent of decisions and AI made 9.2 percent. The most frequent arrangement for methods, at 55.4 percent, had AI put forward options and a person pick one. The tasks judged infeasible without AI showed the same tilt: humans set the goal in 95.4 percent of them.
When tasks hit trouble, people mostly supplied information
The study recorded a difficulty for 588 tasks. In 76 percent of them, a human stepping in got things moving again; in 23 percent, the agent resolved the problem by itself. The human contribution was typically knowledge rather than labour. Supplying context or sharpening requirements accounted for 35.2 percent of cases, and diagnosing a fault or changing approach for 34.7 percent. Hands-on edits were uncommon: 3.2 percent were partial edits and 0.7 percent were complete takeovers.
The authors' concern: review that cannot keep up
The next point is a conditional concern from the authors, not a finding from the problem cases. They worry that every decision could come to rest on a longer chain of agent work than any person can review, and that oversight would then become hard. In the worst case, the team writes, reviewers could only rubber-stamp what they see.
This is where the rise from 11 to 28.5 agent actions per human input matters. The study does not show that review failed. But if each human input covers more agent steps, a count of approvals alone no longer describes how much work a sign-off stands behind. The report adds that many participants ran agents in autonomous modes so that long jobs would not stall on repeated approval prompts. According to The Decoder, that limit reflected convenience rather than a considered view of how much authority AI should hold.
What to ask your team
This is one project's data, not an industry benchmark. It still offers budget-holders a practical way to examine agentic tools in their own engineering and research functions. Which approval gates in our AI-assisted workflows are real decision points, and which merely confirm work already committed? Do we measure how much output would have been infeasible without AI, or only the time saved on work we would have done anyway? Where agents run in autonomous, non-interrupting modes, was that a deliberate risk decision with a named owner, or a default nobody signed off? And who examines the chain of agent steps behind a single human approval, and how long can that chain grow before a reviewer cannot check it meaningfully?
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: The Decoder.
