Guarding the till while autonomous data agents do the digging

Short summary

Autonomous agents querying production databases can cost thousands per run without safeguards. The post demonstrates two guardrails using Google's Data Agent Kit: a BigQuery cost limit that dry-runs queries before execution, and a token budget that pauses for human approval. Concrete Stack Overflow example shows how unoptimized exploratory queries trigger expensive full-column scans.

• •Agents doing exploratory queries on TB-scale tables accumulate thousands in cloud costs per run without guardrails
• •BigQuery charges by bytes read across entire columns, not rows returned—LIMIT and WHERE clauses don't reduce cost
• •Two guardrails solve this: pre-flight query cost checks and per-session token budgets requiring human sign-off

Generated with AI, which can make mistakes.