Generative AI has significantly transformed research workflows. Beyond the direct use of LLMs for programming, writing, and reviewing tasks, agentic frameworks can build on these capabilities to support a broader range of research tasks. Such systems are particularly promising for addressing tedious and time-consuming aspects of  scientific work.

Many agentic frameworks and tools have been proposed for both general and domain-specific tasks. However, building such systems from scratch still requires substantial effort. This motivates the development of a flexible template that allows researchers to easily experiment with different designs or build their own tools.

In this talk, we introduce MAST4Sci (Multi-Agent-System Template for Science), a template that provides core components including support for multiple LLM providers (including SURF AI Hub), planning modules, tool integration, and memory management. The template is designed to be easy to implement, enabling researchers to prototype and evaluate new task-specific agentic frameworks or build practical tools without having to implement the underlying infrastructure from scratch.

To demonstrate its capabilities, we present several implementations built on top of this template, including a metadata-from-dataset tool for research data management (developed in collaboration with SURF) and a meta-analysis tool for agricultural data extraction from scientific literature. These examples illustrate how the template supports the rapid development of diverse agentic applications across research domains.