Data analyst agent

A data analyst agent answers questions by querying databases, running analyses, and producing reports. The core skill: translating natural language questions into SQL (or similar), then interpreting results.

The loop

1. User asks a question
2. Agent understands the data schema (via tool)
3. Agent writes SQL
4. Executes, gets results
5. Interprets results
6. May write more queries for follow-ups
7. Synthesizes final answer with visualizations or tables

Core tools

• list_tables()
• describe_table(table)
• run_query(sql)
• create_chart(data, type)
• explain_plan(sql), for expensive queries

Why schema context matters

LLM can't write correct SQL without knowing table structure. Schema must be in context or discoverable via tools. For large schemas, this is a RAG problem over your DB metadata.

Safety

• Read-only database access (usually)
• Query timeouts (agent doesn't crash the DB)
• Row-level security on the DB side (the agent sees only what the user should)
• Cost caps on long-running queries

Interpretation quality

Numbers without interpretation aren't useful. Agent should:

• State findings in natural language
• Point out surprising numbers
• Flag data quality issues (missing, inconsistent)
• Offer follow-up analyses

When to visualize

Charts for trends, distributions, comparisons. Tables for exact numbers, specific records. Agent picks based on the question.

Common failure modes

• Writes SQL that runs but answers wrong question
• Misinterprets query results (confuses counts with sums)
• Doesn't notice join mistakes
• Confidently answers from bad data

Mitigation: test queries on small samples, sanity-check totals, flag anomalies.