DynaMate2: runtime registration of expert-defined tools for agentic scientific workflow automation

Abstract

Agentic large-language-model systems can coordinate scientific tools, but many implementations remain difficult for domain scientists to extend without modifying the source orchestration code or relying on unconstrained code generation. DynaMate2 is a LangGraph-based multi-agent framework for converting expert-defined Python functions into persistent AI-callable tools. The architecture separates domain execution from LLM supervision: registered tools perform scientific operations, while a supervisor LLM decomposes goals, selects specialist agents, routes inputs, and propagates outputs across steps. DynaMate2 supports: runtime tool registration from inline code, source files, and explicitly requested natural-language specifications; persistent storage of tools, agents, and conversation state; and a web interface for interactive workflow assembly. We demonstrate the framework on a molecular simulation workflow in which a single instruction retrieves a MACE foundation model, builds a NaCl-water configuration, runs an ASE molecular dynamics trajectory, and generates energy and temperature diagnostics. The demonstration illustrates how validated workflow components can be composed into a supervised agentic pipeline without rewriting the framework. DynaMate2 therefore provides a reusable template for extending LLM-based automation to research groups with existing Python workflows, while preserving the need for explicit tool validation, reproducibility logs, and deployment-specific safeguards.