Agent data team

Data team of agents

A number from one agent is a query you cannot see and a pipeline you cannot re-run. Split the work the way a real data team does: a builder makes the pipeline reproducible, a quality checker gates the output, an analyst answers the question with the query attached, and an auditor re-derives every published figure. The auditor is the point, because a confident dashboard on a lossy pipeline is the default failure.

Team

• Builder (data-engineer-role): builds the reproducible pipeline.
• Quality checker (qa-engineer-role, database-testing): asserts on the output tables.
• Analyst (data-scientist-role): answers the question from clean tables.
• Metric auditor: re-derives each headline number by a second path.

Shape: a sequential pipeline with a quality gate before analysis and an audit gate before publishing.

Method

1. Builder writes an idempotent pipeline, not a one-off script. Extract, transform, load with an explicit schema, partition keys, and re-runnable steps (dbt models or an Airflow DAG). It hands off a materialized table plus schema.sql, not a notebook.
2. Quality checker gates on assertions, not eyeballing. Row counts against source, null rates on required columns, primary-key uniqueness, referential integrity, freshness within SLA, and range checks. Output dq-report.md; a failed check blocks analysis.
3. Tie a control total to the system of record. Reconcile revenue or active users to the trusted source within a stated tolerance, for example 0.5%. A pipeline that drops 3% of rows silently must fail here.
4. Analyst answers, and records the query for each number. analysis.md pairs every stat with the exact SQL and the grain (per user, per day). No number ships without a reproducible query.
5. Guard the metric traps. Define the denominator, handle time zones and partial periods, separate event count from entity count, and state the dedup rule. Most wrong dashboards are a denominator or a double-count.
6. Auditor re-derives headline numbers independently. It recomputes each top-line figure from raw source by a different path and signs metric-audit.md only when the two agree within tolerance. It never saw the analyst's queries.
7. Ship against written gates. Release when dq-report.md has no failing check, every published number has a query, and the auditor's re-derivation matches.

Run it

In Claude Code, run builder, checker, analyst, and auditor as sequential subagents over a shared warehouse connection and directory; the orchestrator reads dq-report.md and blocks the analyst subagent if any check failed, and keeps the auditor a separate subagent that never sees the analyst's queries. Port it to CrewAI as a sequential process with the checker as a gating task, to AutoGen as agents where the checker's tool result halts the chain, or to LangGraph as a linear graph with conditional edges on the quality and audit gates.

Signals it works

• Every number in analysis.md has a query beside it that reproduces it.
• The quality gate has actually failed a run and blocked it, not only passed.
• The auditor's independent total matches the analyst's within tolerance.

Boundaries

This runs the build-check-analyze loop; it does not choose which metric matters or own data governance, which the business and a data steward decide. Agents cannot validate against ground truth they cannot query, and a green quality report on biased source data is still biased. Warehouse conventions, SLAs, and tolerance thresholds are yours to set.