Recent breakthroughs in large language models and multimodal architectures have given rise to new “agentic” AI systems capable of performing complex, context-sensitive tasks with minimal human supervision. Their broad appeal spans domains as diverse as healthcare, finance, and scientific research, yet their rapid proliferation also highlights critical challenges in robustness, reproducibility, and real-world reliability. Proprietary silos of development often prevent thorough auditing and restrain the scientific community’s capacity to independently assess or refine these systems—introducing risks of error propagation, bias, and unaccountable decision-making.

In response, open-source methodologies offer a principled path forward. By designing transparent, domain-relevant frameworks for knowledge management and AI-agent support, we can ensure that researchers and practitioners alike retain the agency to drive innovation. Collaboration and shared governance within global networks help democratise access and federate resources, lowering barriers to entry and expanding the collective expertise that fuels method validation and improvement. These bottom-up efforts invite a broader range of stakeholders to participate in iterative design, stress-testing, and validation, ultimately enhancing the systems’ trustworthiness.

Open-source agentic systems are needed to preserve academic freedom and integrity within rapidly advancing AI fields. By learning from early adopters, highlighting the shortcomings of “black-box” solutions through rigorous benchmarks, and discussing strategies to build robust community- and domain-driven ecosystems, we aim to facilitate the creation of frameworks that advance scientific discovery rather than obscuring it. In doing so, we not only promote reproducible science but also uphold the principle that AI’s future—particularly in mission-critical applications—must remain accountable, interdisciplinary, and accessible to all.