Beyond the Chatbot: How Scientific AI Agents Are Overcoming the Limitations of Large Language Models in Biomedical Research

In 2025, Lobentanzer and colleagues catalogued a set of structural limitations that separate conversational large language models (LLMs) from genuine scientific instruments: hallucination, weak domain specificity, an inability to interrogate data directly, poor reproducibility, limited tool access, absent provenance, temporal staleness, incapacity for multi-step scientific workflows, and blindness to multimodal biological information (Lobentanzer et al., 2025). This review does not ask, once more, “what can LLMs do in biomedicine”, a question that has already produced a very large literature of fine-tuned encoders, benchmarks, and narrow applications. Instead, it asks a narrower and more consequential question: which of the nine original limitations have been genuinely addressed by the technological developments of 2025–2026, and which remain open? Screening 130 unique publications retrieved as thematically similar to the Lobentanzer platform paper, we find that the field has moved through at least three architectural generations, LLM as interface, LLM as copilot, LLM as agent, and is now assembling the outline of a fourth, in which foundation models, scientific knowledge layers, data layers, tool layers, and verification layers are stacked into something closer to a scientific operating system than a chatbot. Grounding (retrieval-augmented generation, knowledge graphs, dataset-interrogation agents), tool use, and early multi-agent orchestration have made substantial or partial progress against hallucination, staleness, and the inability to execute analyses. Verification, calibrated uncertainty, causal reasoning, and scientific accountability remain largely unsolved. We propose a diagnostic six-level hierarchy (chatbot → copilot → tool-using assistant → agent → multi-agent system → closed-loop discovery), and argue that scientific intelligence should be understood not as “a larger LLM” but as the sum of a foundation model, scientific grounding, computational agency, verification, and provenance, exercised under continued human oversight.