AI Agents Challenge Human Scientists in Hackathon

François Lanusse, an astrophysicist at the CNRS working at CEA Paris-Saclay, is wondering. "Are AIs smarter than us?" With a group of colleagues, he was beaten in December 2025 during a hackathon, a challenge for programmers, by a team made up of agents, artificial intelligences (AI) capable of making decisions, using different digital tools, developing strategies… autonomously. The challenge was to find the parameters of a model of the evolution of the cosmos capable of explaining a series of observations of the sky influenced by the presence of dark matter.

"The news caused a sensation, even if, in reality, the winning result was the fruit of a collaboration between these agents and humans. But it makes you think!", he remarks.

He is not alone, as the vogue for these artificial agents, which has invaded all sectors of the economy, is also touching science. They are called Kosmos, Co-Scientist, Denario, Sakana, ChemAgent… and they are the representatives of this new family of "self-directed laboratories", "co-scientists", "scientific AIs", "autonomous scientific research", supposed to accelerate research. They consist of several subsystems dedicated to bibliographic research, data analysis, hypothesis formulation, computer code writing, choice of digital tools, trial-and-error repetition, evaluation, writing the final article… and orchestrating all these functions, of course.

Most of these subsystems have a language model as their brain, like those at the heart of ChatGPT, Claude, or Gemini. But agents don't just generate text, they act. On May 19, American companies Google DeepMind and FutureHouse independently published in Nature the promises of their agents, which notably proposed repositioning drugs for pathologies such as leukemia or age-related macular degeneration.