Messy CSV normalization with sentiment tagging is a clean win for a code agent.

The transformation logic — detect delimiter, parse fields, classify sentiment, validate row count — is structurally identical for every row. Once the parsing rules are written, the agent applies them uniformly across all 8,500 records.

Success criteria are concrete: exactly 8,500 rows, five named columns, no data loss, and a three-class sentiment label. The agent can verify all of these programmatically without subjective judgment.

The agent needs only the CSV file and a Python environment with standard libraries (csv, json, pandas, a sentiment model). No external APIs, credentials, or live context are required.

A parsing bug could silently drop or corrupt rows, which would undermine downstream analysis. However, the output is a new file — the original data is untouched — and the 8,500-row validation check makes silent data loss detectable before the result is used.

Delimiter detection and JSON fragment extraction are algorithmic. Sentiment classification on survey text is a well-solved NLP task. No taste, ethics, or relationship context is needed here.