AI agents performed up to 28.5 actions for every human input while building a 744-billion-parameter model, yet humans kept almost all final decisions. A team involving researchers from China's Fudan University analyzed more than 700 task logs from 56 participants who developed Atria Dawn Preview, with AI used in 96.5 percent of tasks. Humans set goals and scope in 93.4 percent of cases, which shows that heavier agent workloads do not yet mean independent development.
How agents and researchers split the work
Over four weeks, the median ratio of agent actions to human inputs rose from 11 to 28.5 as participants delegated more steps to agents. The authors warn against reading this growth as autonomy, since each human decision triggered longer chains of agent work without transferring decision authority. Of 455 completed AI-assisted tasks, 151 were rated infeasible without AI at the same scope and quality, roughly a third of the total. Those tasks were spread across 27 of the 56 participants, so the effect was not limited to a small group of advanced users.
Atria Dawn Preview uses a mixture-of-experts architecture and is designed for research and engineering tasks. Training runs through a pipeline that links each task to a real execution environment, where the model calls tools, produces intermediate results, and is checked against external signals such as tests, metrics, or source evidence. The team reports the lead on five of 16 benchmarks, including web search and cybersecurity, without an overall advantage over competitors. For methods and parameters, the dominant pattern was "AI proposes, human selects" in 55.4 percent of cases, with humans making 85.5 percent of those decisions and AI making 9.2 percent.
The study describes three phases of AI involvement, from research subject to tool for individual tasks to project partner that drafts and adjusts plans within human-set goals. A speculative fourth phase would involve recursive self-improvement, in which stronger models produce stronger successors, but better results on training tasks do not imply better ability to develop a successor. The paper lands amid competing industry signals, with Anthropic reporting a single-digit human share in research direction decisions and OpenAI using GPT-5.6 Sol across its development cycle. Google and DeepMind are testing Dream-RSI to improve search strategy rather than the model itself, while over a thousand employees warn that automation of AI research may be near.
What this means for companies using AI
For companies, agents look less like a way to speed up routine work and more like a way to start work that would otherwise stay untouched. In this project, a third of completed tasks belonged in that category, which matters for engineering, research, and analytics teams with backlogs of complex but well-scoped problems. Smaller firms can use agents to cover missing specialist capacity, while larger organizations can spread execution across teams and keep senior staff focused on goals, methods, and review. Even among the 151 tasks rated infeasible without AI, humans chose the goal 95.4 percent of the time, so scoping skill stays central.
The constraint is oversight and recovery when agent runs grow long. Of 588 tasks with a recorded difficulty, 76 percent moved forward only after human intervention and 23 percent were resolved by the agent alone. Humans helped mainly by adding context or clarifying requirements in 35.2 percent of cases and by diagnosing issues and switching methods in 34.7 percent, while partial edits were 3.2 percent and full takeovers 0.7 percent. When outputs needed revision, AI applied the changes itself in 75.4 percent of cases after feedback. Buyers should therefore check how context is supplied, how failures are diagnosed, and where approval boundaries sit, since participants often ran agents autonomously to avoid interrupting long runs.
The marker to watch is whether agents begin to take a larger share of final decisions on goals and research direction, not just more execution steps. Anthropic chief executive Dario Amodei has called for a speed limit as that share shrinks inside his company. If independent evaluations show agents proposing varied directions and assessing them before results appear, the move toward self-improving systems will look more concrete for business planning.
