Design and Implementation of an Instruction-Driven, Browser-Embedded Business Analysis Agent

Analytical work in supply chain and logistics operations repeats the same pattern: judgment criteria that live in the experience of individual staff, manual calculation in Excel, and report writing and email distribution. This paper designs and implements a browser-embedded business analysis agent with which field staff write their judgment criteria as a natural-language instruction file (Markdown), upload an Excel file, and give a one-line request; the agent then performs analysis, forecasting, report writing, and email distribution. The system consists of a single Chrome extension. A large language model (Gemini) uses function calling to decide the order of tool use and to write the report text, while all numbers are produced by a calculation engine based on a safe expression parser and a forecasting engine based on WebAssembly (Pyodide). The forecasting engine validates linear regression, decision tree, random forest, and XGBoost per group on a time-ordered holdout and selects the best model automatically. Tool-result reduction and conversation-history compaction keep the agent within the per-minute token limit of a free-tier API. Three tasks, namely supply chain risk analysis, workload forecasting per supplier, and delivery schedule notification, were performed without any program change by replacing only the instruction file and the Excel file. On 845 rows of inbound history from five suppliers, automatic selection achieved a mean RMSE of 38.26, 9.3% lower than the best single fixed algorithm, with errors of 4.6–7.6% of the mean inbound volume. Forecasting completed in 21.7 s on the first run and 4.4 s thereafter, and input tokens per report decreased by 80.7%. The system presents a practical configuration that lets field staff record their own judgment criteria as documents and apply them to recurring work without servers or paid tools.