Agentic AI: A Tool-Guided Approach to Achieving Precision in Engineering Numerical Operations

This paper introduces an agentic AI architecture that augments Large Language Models (LLMs) with tool-guided workflows to achieve consistent and dependable mathematical extraction and computation. By guiding LLMs through structured, agentic workflows, the proposed system addresses their inherent underperformance in mathematical operations and numerical consistency, despite strong contextual understanding. The system decomposes tasks into a sequence of modular, autonomous agents—each assigned a distinct function—ensuring high-precision outcomes in data-driven engineering applications. The architecture operates through a well-defined process: a Document Agent ingests and preprocesses unstructured inputs (e.g., PDF engineering reports); an Extraction Agent utilizes an LLM to identify and extract structured information (e.g., acid volumes, fluid stages, depth ranges); a Validation Agent performs schema-based validation to ensure data integrity; a Computation Agent executes deterministic calculations on validated data (e.g., summing acid volumes, determining total Viscoelastic Diverting Acid volumes, and calculating depth intervals); and a Reporting Agent formats and outputs the results as structured tables or JSON. The framework was tested using acid job reports, successfully parsing, validating, and processing examples such as "PUMP 43.53 BBLS PRE FLUSH" and "15% HCl MAIN ACID 21.3 BBLS". This demonstrated the robustness of the agentic workflow in handling technical language and numerical operations, even in cases prone to LLM misinterpretations like fluid type differentiation or concentration calculations. Key observations include improved consistency due to schema enforcement, significant mitigation of common LLM errors (e.g., incorrect unit interpretation, misclassification of fluid types, erroneous percentage calculations), and accurate results even in ambiguous cases. By eliminating numerical hallucinations and ensuring accurate data aggregation, the framework consistently outperforms raw LLM approaches, which often yield inconsistent or incorrect calculations despite prompt tuning. This guided, tool-augmented approach enhances LLMs’ contextual understanding while delegating mathematical fidelity to deterministic tools, providing a robust method for deriving accurate engineering insights from complex data. The framework is generalizable to other domains such as cementing, fracturing, and economic evaluation, bridging natural language understanding with deterministic computation for scalable, trustworthy LLM deployment in real-world energy-sector applications.