Toward autonomous science with agentic artificial intelligence

Biomedical AI has rapidly transitioned from simple pattern recognition in medical imaging to generative models like AlphaFold that predict protein structures. While these earlier tools were powerful, they operated within fixed boundaries defined by human researchers. The emergence of agentic AI—systems capable of reasoning, planning, and executing multi-step scientific workflows—marks a fundamental shift toward autonomous science. Three recent systems, Co-Scientist, Robin, and Biomni, demonstrate this new capability by autonomously generating hypotheses, designing experiments, and interpreting data through iterative feedback loops.

The field is currently exploring diverse paths to building an 'artificial scientist.' Co-Scientist utilizes a multi-agent 'tournament' architecture, where competing agents generate and critique hypotheses, using test-time compute to refine proposals before human review. Robin takes a similar multi-agent approach but integrates experimental data analysis directly into the loop, allowing it to propose and refine therapeutic candidates based on wet-lab results. In contrast, Biomni acts as a general-purpose agent that uses code-as-action to interact with specialized databases and software, dynamically composing workflows without predefined templates. These systems represent a move away from single-pass predictions toward iterative, reasoning-based discovery.

Despite their promise, these systems face significant hurdles. Hallucinations remain a primary concern, as AI-generated hypotheses may appear plausible but lack a basis in biological reality. Furthermore, the 'ground truth' in scientific discovery is often elusive, making it difficult to evaluate these systems without long-term experimental validation. There is also a risk of bias amplification, where models trained on existing literature reinforce established orthodoxies rather than exploring novel, unconventional ideas.

Looking forward, the most transformative potential lies in the 'autonomous laboratory'—a closed-loop system where AI reasoning is directly coupled to robotic hardware. This would remove the human latency between hypothesis and experiment, allowing for rapid, continuous iteration. However, realizing this vision requires not only better AI reasoning but also standardized data infrastructure, open orchestration frameworks, and rigorous governance to manage dual-use risks and ensure scientific reproducibility.