Steps towards agentic AI

Current large language models optimised for fluency can hallucinate and only offer the appearance of reasoning. This work proposes research into using TAG (Tool-Augmented Generation) and RAG agents to serve as verifiable, deterministic executors of scientific software, while higher-level coordinators handle task decomposition and synthesis, thereby replacing statistical ambiguity with exact, engineer-like factual grounding. The concept of Custodian Transfer Contracts is introduced, aimed at capturing domain expertise and transition software interaction from human-read manuals to TAG-wrapped computational libraries.

This requires a scientific library that the TAG-agents can call. Therefore, this talk is organized into three parts:

• General framework is the agentic architecture, the role of TAG/RAG agents, and Custodian Transfer Contracts.
• A scientific library that can be used by the TAG-agents, comprised of two parts:
  • Scalars and Applications – polymorphic, move-semantic compliant collection of advanced scalars (CPX, SU2, CL(1,3)), high-speed algorithms (stochastic transformer, non-abelian Gauss-Jordan, flash vertexing/helix fitting), and dedicated Fourier tools (FoxLima, DXF, Weyl-Wigner).
  • Neurosoftware expertise – derived in work with Extreme Learning Machines (ELM), Random Fourier Features (RFF), MLPfit, NARX, and a Bayesian 3D U-net.

This scientific library with its varied chapters is an ideal testbed for training TAG-agents, its tools collectively addressing actionable demands of Scientific-AI.