Data Analyst Agent — Tool-Calling with Tracing and Human-in-the-Loop

An agent that takes a question (in Arabic or English) about an orders dataset, plans its steps, and calls tools. It checks data quality first, cleans duplicates and missing values when it finds them, aggregates, plots a chart, extracts insights, drafts a report, and asks for human approval before saving any file. Every step is recorded in a timed trace. The guardrails are implemented and covered by 8 passing tests: a step budget, an approval gate (rejecting writes nothing), tool-error handling, argument validation, and a check that the reported numbers match an independent calculation. The dataset is synthetic and the default planner is rule-based; an LLM-based planner interface is included but has not been tested yet.