Agent

A comprehensive review from Cell envisioning 'AI scientists' as systems capable of skeptical learning and reasoning that empower biomedical research through collaborative agents integrating AI models and biomedical tools with expert human oversight. Covers the landscape of AI agents for biomedical discovery, including virtual laboratories, autonomous experimentation, and human-AI collaboration frameworks.

We envision 'AI scientists' as systems capable of skeptical learning and reasoning that empower biomedical research through collaborative agents that integrate AI models and biomedical tools with expert human oversight. This review covers the landscape of AI agents for biomedical discovery, including virtual laboratories, autonomous experimentation, and human-AI collaboration frameworks.

A comprehensive review from Nature Reviews Cancer examining how AI agents (beyond traditional ML classifiers) are transforming cancer research and oncology. Covers LLM-powered agents for clinical decision support, drug discovery, treatment planning, and patient care, including agentic systems capable of logical reasoning, multi-step planning, and tool use in oncology contexts.

Since 2022, artificial intelligence (AI) methods have progressed far beyond their established capabilities of data classification and prediction. Large language models (LLMs) can perform logical reasoning, multi-step planning, and tool use, enabling a new paradigm of AI agents for cancer research and oncology. This review examines how AI agents are transforming clinical decision support, drug discovery, treatment planning, and patient care in oncology.

AI-MARRVEL (AIM) is a knowledge-driven AI system for diagnosing Mendelian disorders that uses a random-forest machine-learning classifier trained on over 3.5 million variants from thousands of diagnosed cases. It incorporates expert-engineered features to recapitulate the complex decision-making processes in molecular diagnosis. AIM doubled the rate of accurate genetic diagnosis across three independent real-world cohorts compared to benchmarked methods. Its confidence metric achieved 98% precision and identified 57% of diagnosable cases from 871 unsolved cases. AIM also demonstrated potential for novel disease gene discovery.

Diagnosing genetic disorders requires extensive manual curation and interpretation of candidate variants, a labor-intensive task even for trained geneticists. Although artificial intelligence (AI) shows promise in aiding these diagnoses, existing AI tools have only achieved moderate success for primary diagnosis. AI-MARRVEL (AIM) uses a random-forest machine-learning classifier trained on over 3.5 million variants from thousands of diagnosed cases. AIM additionally incorporates expert-engineered features into training to recapitulate the intricate decision-making processes in molecular diagnosis. AIM improved the rate of accurate genetic diagnosis, doubling the number of solved cases as compared with benchmarked methods, across three distinct real-world cohorts. AIM achieved a precision rate of 98% and identified 57% of diagnosable cases out of a collection of 871 cases. AIM demonstrated potential for novel disease gene discovery by correctly predicting two newly reported disease genes from the Undiagnosed Diseases Network.

BioMedAgent is a self-evolving LLM multi-agent framework that learns to use diverse bioinformatics tools and chain them into executable workflows through interactive exploration and memory retrieval algorithms. It allows biomedical users to initiate tasks using natural language, without requiring computational expertise. Evaluated on BioMed-AQA benchmark (327 biomedical data tasks), BioMedAgent achieved a 77% success rate, surpassing other LLM agents, and generalized robustly to the external BixBench dataset. Beyond benchmarks, it autonomously performs cross-omics analysis, machine-learning modelling and pathology image segmentation.

Artificial intelligence agents are emerging as powerful applications of large language models (LLMs), automating complex tasks and enabling scientific data exploration. However, their use in biomedical data analysis remains limited by the difficulty of handling specialized tools and multistep reasoning. Here we introduce BioMedAgent, a self-evolving LLM multi-agent framework, which learns to use diverse bioinformatics tools and chain them into executable workflows through interactive exploration and memory retrieval algorithms. It allows biomedical users to initiate tasks using natural language, without requiring computational expertise. Evaluated on our newly released BioMed-AQA benchmark comprising 327 biomedical data tasks, BioMedAgent achieved a 77% success rate, surpassing other LLM agents, and generalized robustly to the external BixBench dataset. Beyond benchmarks, it autonomously performs cross-omics analysis, machine-learning modelling and pathology image segmentation, highlighting its potential to advance biomedical research and extend to other scientific domains requiring complex tool integration and multistep reasoning.

Biomni is a general-purpose biomedical AI agent designed to autonomously execute a wide spectrum of research tasks across diverse biomedical subfields. It employs an action discovery agent to mine tools, databases, and protocols from tens of thousands of publications across 25 biomedical domains, creating the first unified agentic environment (Biomni-E1). Its generalist agentic architecture (Biomni-A1) integrates LLM reasoning with retrieval-augmented planning and code-based execution, dynamically composing complex workflows without predefined templates. Systematic benchmarking demonstrates strong zero-shot generalization across heterogeneous tasks including causal gene prioritization, drug repurposing, rare disease diagnosis, microbiome analysis, and molecular cloning.

Biomedical research underpins progress in our understanding of human health and disease, drug discovery, and clinical care. However, with the growth of complex lab experiments, large datasets, many analytical tools, and expansive literature, biomedical research is increasingly constrained by repetitive and fragmented workflows that slow discovery and limit innovation. Here, we introduce Biomni, a general-purpose biomedical AI agent designed to autonomously execute a wide spectrum of research tasks across diverse biomedical subfields. To systematically map the biomedical action space, Biomni first employs an action discovery agent to create the first unified agentic environment, mining essential tools, databases, and protocols from tens of thousands of publications across 25 biomedical domains. Built on this foundation, Biomni features a generalist agentic architecture that integrates LLM reasoning with retrieval-augmented planning and code-based execution, enabling it to dynamically compose and carry out complex biomedical workflows entirely without relying on predefined templates or rigid task flows. Systematic benchmarking demonstrates that Biomni achieves strong generalization across heterogeneous biomedical tasks including causal gene prioritization, drug repurposing, rare disease diagnosis, microbiome analysis, and molecular cloning without any task-specific prompt tuning. Real-world case studies further showcase Biomni's ability to interpret complex, multi-modal biomedical datasets and autonomously generate experimentally testable protocols.

ChemCrow is an LLM-based agent system that augments large language models with 13 expert-designed chemistry tools, enabling autonomous chemical reasoning and experiment design. It integrates tools for organic synthesis, drug discovery, and materials design, allowing the agent to plan syntheses, analyze chemical properties, and execute complex chemical tasks through natural language interaction.

Large language models (LLMs) have shown strong performance in tasks across domains but struggle with chemistry-related problems. These models also lack access to external knowledge sources, limiting their usefulness in scientific applications. Here we introduce ChemCrow, an LLM-based agent that integrates 13 expert-designed chemistry tools, enabling autonomous chemical reasoning and experiment design. ChemCrow successfully plans syntheses, analyzes chemical properties, and executes complex chemical tasks through natural language interaction.

Co-Scientist is a multi-agent AI system built on Gemini for structured scientific thinking and hypothesis generation. It aims to help scientists discover new original knowledge by formulating demonstrably novel research hypotheses for experimental validation, conditioned on research objectives and prior scientific evidence.

Scientific discovery is driven by scientists generating novel hypotheses for complex problems that undergo rigorous experimental validation. To augment this process, we introduce Co-Scientist, a multi-agent AI system built on Gemini for structured scientific thinking and hypothesis generation. Co-Scientist aims to help scientists discover new original knowledge. Conditioned on their research objectives and prior scientific evidence, it formulates demonstrably novel research hypotheses for experimental validation.

CRISPR-GPT is an LLM-based agent system for automating gene-editing experiments. It leverages large language models to guide researchers through the entire CRISPR experiment workflow, including guide RNA design, off-target prediction, experimental protocol generation, and result interpretation, making gene-editing more accessible to non-expert researchers.

Performing effective gene-editing experiments requires a deep understanding of both the CRISPR technology and the biological system involved. Meanwhile, despite their versatility and promise, large language models have not been fully leveraged for automated experimental design in molecular biology. Here we present CRISPR-GPT, an LLM-based agent system for automating gene-editing experiments across the entire CRISPR workflow, including guide RNA design, off-target prediction, experimental protocol generation, and result interpretation.

DeepRare is a multi-agent system for rare disease differential diagnosis that integrates large language models with structured medical knowledge to provide traceable reasoning. It aims to reduce the diagnostic odyssey for rare disease patients by leveraging AI agents to analyze clinical phenotypes, genomic data, and medical literature in a transparent, interpretable manner.

Rare diseases affect more than 300 million people worldwide, yet timely and accurate diagnosis remains an urgent challenge. Patients often endure a prolonged 'diagnostic odyssey' exceeding 5 years, marked by repeated referrals, misdiagnoses and unnecessary interventions, leading to delayed treatment and substantial emotional and economic burden. Here we present DeepRare, a multi-agent system for rare disease differential diagnosis decision support with traceable reasoning.

Hermes Agent is an open-source autonomous AI agent built by Nous Research with a built-in closed learning loop — it creates skills from experience, improves them during use, persists knowledge, searches past conversations, and builds a user model across sessions. Supports 14 messaging platforms (Telegram, Discord, Slack, WhatsApp, Signal, etc.) and 6 execution backends (local, Docker, SSH, Modal, Singularity). Features include persistent memory (FTS5 + Honcho), automated skill creation, cron scheduling, parallel sub-agent delegation, full browser automation, MCP integration, and model-agnostic provider switching via CLI. MIT licensed, first released February 2026.

A natural-language interface that enables large language models (LLMs) to perform end-to-end variant interpretation for Mendelian diseases via structured tool access (MCP). MARRVEL-MCP equips LLMs with 44 tools spanning gene and variant utilities, pathogenicity databases, phenotype resources, expression atlases, ortholog data, and literature APIs. Without hard-coded workflows, LLMs infer which tools to invoke and in what sequence, performing named-entity recognition, identifier normalization, and multi-database synthesis from clinical queries. A 20B-parameter model achieved 94% pass rate on 100 expert-curated questions (vs 41% without MARRVEL-MCP), approaching state-of-the-art proprietary performance. Establishes context engineering as a core principle for biomedical AI.

Variant interpretation in rare diseases requires navigating multiple genomic databases, each with strict input formats, while synthesizing heterogeneous evidence. To address these usability barriers, we developed MARRVEL-MCP, a natural-language interface that enables large language models (LLMs) to perform end-to-end variant interpretation via structured tool access. MARRVEL-MCP equips LLMs with 44 tools spanning gene and variant utilities, pathogenicity databases, phenotype resources, expression atlases, ortholog data, and literature APIs. Without hard-coded workflows, LLMs infer which tools to invoke and in what sequence, performing named-entity recognition, identifier normalization, and multi-database synthesis from clinical queries. Using 100 expert-curated questions, lightweight models (3B-20B parameters) with MARRVEL-MCP matched or outperformed larger models without tool access. A 20B-parameter model achieved a 94% pass rate, versus 41% without MARRVEL-MCP, approaching state-of-the-art proprietary performance. These findings establish context engineering as a core principle for biomedical AI and support scalable integration of LLMs with curated genomic resources.

MIRA is an autonomous AI agent powered by GPT-4o (T=0.01) with o1-preview for structured reasoning, operating within a sandboxed HL7 FHIR-based EHR environment. It navigates 85,000+ clinical decision options across 8 emergency department diagnoses, using 11 FHIR-compliant tools (PatientHistory, PhysicalExam, Lab/Urine/Microbiology/Radiology requests, Medication/Procedure ordering, Plan, Admission). Evaluated on 574 real MIMIC-IV patient cases, MIRA outperformed two independent physician cohorts in diagnostic accuracy, guideline-concordant treatment, medication safety, and appropriate admission decisions. All tool parameter validity is enforced through token masking, making hallucination of non-existent options programmatically impossible.

Large language models (LLMs) show great potential for clinical decision-making, yet most applications remain narrow, task-specific chat tools rather than systems integrated into clinical workflows. However, building physician copilots will require models that operate within the electronic health record (EHR), with governed access to patient data and the ability to initiate permitted EHR actions within defined safety constraints. Here we show that MIRA (Medical Intelligence for Reasoning and Action), an autonomous artificial intelligence agent operating in a sandboxed EHR environment, can navigate a large clinical action space to obtain patient histories; order and interpret laboratory, imaging and microbiology tests; generate differential diagnoses; and formulate treatment plans such as prescribing medications, scheduling surgical procedures and planning admissions. In simulations on real patient cases spanning multiple diagnoses, MIRA outperformed physicians in diagnostic accuracy and made guideline-concordant, medication-safe and appropriate admission decisions.

Open-source, local-first autonomous AI assistant: file/shell access, browser automation, skills/plugins, and optional chat-app bridges; model-agnostic (e.g. Claude, OpenAI, local/Ollama) with your own API keys.

Robin is the first multi-agent system capable of fully automating both hypothesis generation and data analysis for experimental biology. By integrating literature search agents with data analysis agents, Robin can generate testable hypotheses from literature and design experiments to validate them, automating the entire scientific discovery cycle for biological research.

Scientific discovery is driven by the iterative process of observation, hypothesis generation, experimentation, and data analysis. Despite recent advancements in applying artificial intelligence to biology, no system has yet automated all these stages. Here, we introduce Robin, the first multi-agent system capable of fully automating both hypothesis generation and data analysis for experimental biology. By integrating literature search agents with data analysis agents, Robin can generate testable hypotheses from literature and design experiments to validate them.

SciSciGPT is an open-source, prototype AI collaborator that uses the domain of science of science as a testbed to explore LLM-powered scientific collaboration. It assists researchers in analyzing scientific literature, identifying research trends, generating hypotheses about scientific dynamics, and facilitating human-AI collaborative research in metascience and scientometrics.

We introduce SciSciGPT, an open-source, prototype artificial intelligence (AI) collaborator that uses the domain of science of science as a testbed to explore the potential of large language model-powered scientific collaboration. SciSciGPT assists researchers in analyzing scientific literature, identifying research trends, generating hypotheses about scientific dynamics, and facilitating human-AI collaborative research in metascience and scientometrics.

SciToolAgent is a knowledge-graph-driven scientific agent that integrates multiple computational tools for scientific research. It leverages a knowledge graph to understand tool capabilities and relationships, enabling automated multi-tool workflow composition for complex scientific tasks across domains including bioinformatics, cheminformatics, and materials science.

Scientific research increasingly relies on specialized computational tools, yet effectively utilizing these tools requires substantial domain expertise. While large language models show promise in tool use, they struggle with complex multi-tool workflows. Here we introduce SciToolAgent, a knowledge-graph-driven scientific agent that integrates multiple computational tools for scientific research. By leveraging a knowledge graph to understand tool capabilities and relationships, SciToolAgent enables automated multi-tool workflow composition for complex scientific tasks.

SPARK (System of Pathology Agents for Research and Knowledge) is a foundational agentic AI framework that uses language as a universal interface to autonomously generate biologically driven concepts for tumor analysis. It functions as a pathology 'brain' — an interconnected system of AI agents that autonomously reason, generate and implement biologically meaningful hypotheses as analytical tools without additional model training. SPARK uses four linked modules: idea generation (OpenAI o1), idea refinement, parameter coding (Claude Sonnet 3.5), and parameter verification. Evaluated across 18 patient cohorts spanning 5 cancer types and >5,400 patients, SPARK produced clinically and biologically relevant concepts correlated with prognosis, pathological variables, and predictive biomarkers, including patterns of tumor progression inferred from static images.

Artificial intelligence has advanced cancer pathology, but many systems still depend on hand-crafted features, are hard to explain and rely on fragmented workflows. We introduce SPARK (System of Pathology Agents for Research and Knowledge), a foundational agentic artificial intelligence approach that uses language as a universal interface to autonomously generate biologically driven concepts for tumor analysis. SPARK turns biological ideas into analytical tools and works directly with complex pathology data without extra model training. We evaluated SPARK across 18 patient cohorts spanning five cancer types (lung adenocarcinoma, lung squamous cell carcinoma, colorectal cancer, breast cancer and oropharyngeal squamous cell carcinoma) and more than 5,400 patients with available histopathology images and clinical/follow-up information, in both prognostic and predictive settings and on a well characterized spatial biology breast cancer dataset (n=625). We found that SPARK produced clinically and biologically relevant concepts correlated with prognosis, known pathological variables and predictive biomarkers, including patterns of tumor progression and temporal change inferred from static images. A dedicated module allows for human interaction with SPARK. All code, parameters and results are openly released.

The Virtual Lab is an AI agent framework that uses LLM-powered researchers in a simulated laboratory environment to collaboratively design and test scientific hypotheses. It was demonstrated by successfully designing new SARS-CoV-2 nanobodies, with AI agents specializing in different scientific roles working together to propose, evaluate, and refine experimental designs.

Science frequently benefits from teams of interdisciplinary researchers, but many scientists do not have easy access to experts from multiple fields. Although large language models (LLMs) have shown an impressive ability to aid researchers across diverse domains, their uses have been largely limited to answering specific scientific questions rather than performing open-ended research. Here we expand the capabilities of LLMs for science by introducing the Virtual Lab, a framework where LLM-powered AI agents collaborate in a simulated laboratory to design and test scientific hypotheses. The Virtual Lab successfully designed new SARS-CoV-2 nanobodies, demonstrating the potential of multi-agent AI systems for open-ended scientific discovery.