AI agents in drug discovery: A review of evolution, applications, and future directions

Artificial intelligence (AI) agents represent a paradigm shift in pharmaceutical research, moving the field from narrow drug-protein affinity modeling toward systems-biology-level evaluation in which autonomous, multi-domain agents combine pattern recognition with symbolic reasoning, knowledge graphs, and regulatory intelligence. This review traces the evolution of AI agents in drug discovery across four eras - database systems (1990-2012), machine learning (2012-2022), foundation learning tools (2022-2023), and autonomous agents (2023-present) - and analyzes breakthrough systems including AlphaEvolve, Google's AI Co-scientist, DrugAgent, Boltz-1/Boltz-2, and Isomorphic Labs' clinical programs, reporting industry-disclosed estimates of 25%-30% improvements in Phase I success rates and 30%-40% reductions in preclinical costs together with their statistical limitations. We present a taxonomy of next-generation architectures spanning foundation model-based agents, autonomous multi-agent ecosystems with explicit coordination protocols (consensus voting, debate, hierarchical orchestration), and specialized systems for target discovery, molecular design, and clinical optimization, situating them within knowledge-graph and neuro-symbolic reasoning (PrimeKG, Hetionet, AnyBURL; Hit@K, MRR, AUROC) and the emerging Internet of Agents. We introduce an enhanced Autonomy-Trust Framework that links four levels of autonomous capability to corresponding trust infrastructure and concrete validation strategies, including +Masking and +LLMEval ablations for Levels 2 and 3. Applications in biomarker discovery and precision medicine are examined alongside challenges in validation, data quality, regulatory compliance, and ethics. Current evidence positions AI agents as transformative tools, with Level 2 collaborative agents becoming mainstream while Level 3 autonomous specialists emerge in focused domains.