Defense in depth instead of a single lock
On the agent that queries an analytics database in natural language, three independent layers filter every generated query: expected-schema validation, a whitelist of allowed tables, and a strictly read-only database role. If one layer fails to block a problematic query, the other two stay active — never a single point of failure.
Automation prepares, it doesn't decide alone
Even with automated quality control scoring and filtering every output, final approval before publishing stays human. The system's job is to reduce the volume of trivial decisions to make — not to remove the decision itself on anything that puts the brand on the line.
Observing, not just logging
A standard application log tells you an error happened. Observability built for an agentic pipeline tells you where, on which node, how long it took and how many tokens it burned — and lets you set alert thresholds before a cost or quality drift becomes a real problem. On the systems we run, that tracing goes through Arize Phoenix, node by node.