Tasks in the study where AI was involved: 96.5% (Source: The Decoder (September 28, 2026), reporting on the Atria research team's study)
Agents did more of the work, but not more of the deciding
A team affiliated with China's Fudan University set out to answer a narrow question about its own project: as AI agents took on more of the execution, did the humans running the show actually give up control over what got decided? The researchers pulled together a large batch of records — more than 700 task logs in all — covering 56 people who built Atria Dawn Preview, a 744-billion-parameter agentic language model, and matched each one against the corresponding logs from the AI agents assigned to the same work. Across the project, AI touched nearly every task logged, but the study's own framing pushes back on reading that as agents running the show.
The split between agent and human activity shifted as the weeks went on. Early in the project, a typical task involved a median of 11 agent actions for every human input; by the fourth week that ratio had climbed to roughly 28.5 agent actions per human input. The researchers cautioned against reading that shift as agents making more independent judgment calls — a single human instruction was simply triggering a longer chain of downstream agent steps, not a bigger share of agent decision-making.
Work that would not have happened at all
The more striking finding sits away from the autonomy question. Participants were asked whether they could have completed their portion of a task, at the same scope and quality, without AI. About a third of completed AI-assisted tasks were rated infeasible without it — spread across 27 of the 56 participants, or roughly 48% of the group, so the finding wasn't concentrated among a handful of heavy users. In those cases, AI was not accelerating work already planned. It was the reason the work existed at all.
Direction stayed with people
On the decisions that set direction — what to build, which method to use, when something counted as finished — humans kept authority even inside the AI-dependent tasks. The most common pattern for choosing a method was AI proposing options and a human selecting among them. Final calls on goals and scope stayed almost entirely with people, including in the subset of tasks the team had already flagged as impossible without AI.
Oversight strains when problems appear
The pattern held when things broke. Most recorded difficulties were resolved through human intervention, but that intervention was almost always information — adding context or diagnosing what had gone wrong — rather than a person doing the work directly; full takeovers were rare. The remaining share of problems, the agents resolved on their own. The team's own caution is worth carrying into any organization running similar workflows: once a chain of agent actions grows longer than a person can realistically review, oversight risks becoming, in the researchers' words, reviewers who can only 'rubber-stamp what they see.' Some participants had already started running agents in autonomous modes specifically to avoid interrupting long jobs for approval — a boundary set for convenience, not as a deliberate call on how much authority to hand over.
What to ask your team
This is one project's data, not an industry benchmark, but it gives budget-holders a concrete way to check how agentic tools are actually being used inside their own engineering and research functions. Worth asking: How many approval gates in our AI-assisted workflows are genuine decision points versus click-through confirmations of work already committed? Do we track how much of our AI-assisted output would have been infeasible without it, or are we only measuring speed-up on work we'd have done anyway? Where teams have moved agents into autonomous, non-interrupting modes, was that a deliberate risk decision with an owner, or a convenience default nobody signed off on? And who reviews the chain of agent steps behind a single human approval — and how long is that chain allowed to get before a reviewer can no longer meaningfully check it?
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.
