How developers can cap costs when AI agents query large data warehouses autonomously

Autonomous AI agents can answer complex business questions by querying databases and iterating through trial-and-error, but this loop can become extremely expensive on large platforms like BigQuery. Unlike human analysts who scope queries carefully, agents may run unoptimized scans across entire columns, with a single 100TB query potentially costing over $600 at standard pricing. A developer using Google's Antigravity SDK and the Data Agent Kit built a cost-safe agent with two guardrails: one that dry-runs every query before execution to block oversized scans, and another that pauses the agent for human approval once token spending crosses a set budget. The post highlights that common assumptions — such as LIMIT clauses reducing scan costs — are false in columnar databases like BigQuery, where charges are based on bytes read rather than rows returned. The guardrails aim to make autonomous data agents practical in production environments without exposing organizations to runaway infrastructure costs.