AI Agents Execute, Humans Set Goals

Researchers who built Atria Dawn Preview, a model with 744 billion parameters, analyzed hundreds of task logs. Agents perform most of the execution and enable unprecedented work, but the choices of method, objective, and scope remain largely human. This organization complicates supervision as the industry explores the automation of its R&D.

Supervision Under Scrutiny and Automation in Debate

When the sequence of actions performed by agents becomes too complex for a person to thoroughly examine, ensuring effective supervision becomes challenging. In the worst-case scenario, humans merely approve what is submitted to them. Many participants have used autonomous modes to avoid repeated approval interruptions, a boundary often set for convenience rather than after an explicit choice regarding the authority granted to AI. Meanwhile, major players are experimenting with more automated approaches. Anthropic believes it is possible sooner than expected for an AI to design its successor, and its CEO Dario Amodei calls for a speed limit for the sector. According to the company, humans would only take part in single-digit percentages of research direction decisions. OpenAI uses GPT-5.6 Sol throughout its development cycle. Google and DeepMind allow agents to explore strategies via Dream-RSI, with a scope limited to improving research strategy rather than the model itself. Over a thousand employees from major AI companies recently raised concerns about a potential shift towards research automation. Researchers from Princeton and the UK’s AI Safety Institute conclude that current models can conduct research engineering but fail on significant judgment decisions.

Who Decides: AI Proposes, Humans Decide in Most Cases

In choices of methods and parameters, the most common pattern is where AI proposes and humans select, observed in 55.4% of situations. Overall, humans made 85.5% of decisions regarding methods and parameters, compared to 9.2% for AI. For objectives and scope, the final decision was human in 93.4% of cases. The proportion of proposals coming from AI varies by type of decision, ranging from 17% to 55%, but its share in final decisions remains in single digits. Despite variations in the origin of options, humans retain the majority of final choices. Even for the 151 tasks deemed unachievable without AI, the definition of the objective was chosen by humans in 95.4% of cases.

What Assistance Changes: More Agentic Execution, Central Human Judgment

AI was mobilized in 96.5% of the tasks studied. The median ratio of agent actions to human inputs increased from 11 to 28.5 in four weeks. The authors of the study caution against interpreting these results as an increase in autonomy: each human intervention primarily led to a multiplication of steps performed by agents. Among the 455 tasks that benefited from assistance, 151 were considered impossible to accomplish without AI, with these cases distributed across 27 of the 56 participants, indicating a diffuse effect. In these cases, AI did not simply enable faster execution of an existing task; it allowed for the execution of tasks that could not have been completed otherwise. In the face of difficulties, human intervention facilitated progress in 76% of the 588 reported cases, while the agent resolved 23% on its own. Human assistance took the form of information contributions: adding context and clarifying requirements in 35.2%, diagnosing and changing methods in 34.7%. Direct interventions were rare, with 3.2% of partial modifications and 0.7% of complete takeovers. When a revision was necessary, AI itself made 75.4% of corrections after feedback, indicating that the bottleneck lies on the judgment side rather than in execution.

The Atria Project: An MoE Model and a Pipeline Verified by External Signals

A project led by a team including researchers from Fudan University served as the basis for this work. Over 700 task logs from 56 participants, as well as agent logs, were analyzed. The work focused on Atria Dawn Preview, an agentic language model based on a mixture of experts architecture totaling 744 billion parameters, designed for applications in research and engineering. This model was trained using a pipeline connected to a real execution environment, allowing for the use of tools, the generation of intermediate results, and validation through external signals, such as tests, metrics, or source evidence. According to the team, the model ranks first on five of the sixteen evaluated benchmarks, including web search and cybersecurity, although it does not claim overall superiority. The researchers describe a three-phase trajectory of the role of AI, moving from a research subject to a tool, and then to a project partner, the latter developing and adjusting plans within objectives set by humans. They mention a hypothetical fourth phase of recursive self-improvement, where more powerful models would produce stronger successors.