Task logs from the Atria Dawn build place the bottleneck at human judgment

The step ratio divides agent actions by human inputs. It climbs whenever one instruction fans out into a longer run ^[6]. Over the four weeks the median climbed about 2.6 times ^[3]. Authority over outcomes stayed with people. AI's share of proposals ranged from 17 to 55 percent depending on decision type, while its share of final decisions stayed in single digits ^[12]. For methods and parameters, the most common pattern was "AI proposes, human selects," at 55.4 percent, and AI made 9.2 percent of those calls ^[9][10].

The new-work figure comes to about 33 percent of completed AI-assisted tasks ^[1]. The other 304 tasks, about two-thirds, were ones participants said they could have finished at the same scope and quality without AI ^[2]. So in these logs agents mostly helped with work that would have happened anyway, and the added scope is one task in three ^[2]. The feasibility rating is each participant's own counterfactual, collected by a team studying its own project ^[7][1].

When a task stalled, people mostly supplied information. Adding context or clarifying requirements accounted for 35.2 percent of cases, and diagnosing issues or switching methods for 34.7 percent, about 70 percent together ^[14][4]. Partial edits and full takeovers came to 3.9 percent combined ^[5]. After human feedback, the agent made the revision itself 75.4 percent of the time ^[15]. The Decoder's summary of the study puts the bottleneck at human judgment ^[16].

I think the right plan on this evidence is to design agent workflows around decision throughput. That means spending effort on the handoff where a person gives context or picks among proposals. It also means making agent output reviewable at the length it now runs. The logs record a team that kept its decisions, so they cannot say how the work would have gone had it handed them over ^[1].

The numbers also belong to one kind of workload. Atria Dawn Preview is a 744-billion-parameter mixture-of-experts model built for research and engineering tasks ^[2]. Its training pipeline tied each task to a real execution environment and checked results against tests, metrics or source evidence ^[3]. For the intervention mix to carry over, a team needs tasks with an external signal the agent can fail against. It also needs people who can diagnose the failure quickly. Teams whose tasks have no test to fail should expect a different mix.

The weak point was the approval boundary. Many participants ran agents in autonomous modes so long runs would not stop for constant approvals ^[18]. The team says convenience drew that line, with no deliberate choice about how much authority AI should hold ^[18]. It warns that when each decision rests on a longer chain of agent work than any person can review, humans can end up as reviewers who only rubber-stamp what they see ^[17]. AI was used in 96.5 percent of the tasks reviewed ^[4].