Nov 12, 2025

Stale Context Is Worse Than No Context

An agent with no schema information asks a clarifying question. An agent with stale schema information doesn't know it's wrong, so it doesn't ask — it just answers, confidently, on top of a table that changed last week. That's the more dangerous failure mode, and it's the one most teams aren't watching for, because nothing throws an error when it happens.

It's also a mechanical, fixable problem. Snowflake's INFORMATION_SCHEMA has no replication lag; ACCOUNT_USAGE, which most tooling defaults to, can run anywhere from 45 minutes to a few hours behind. Poll the wrong one on a schedule that feels fast, and you've built something that is confidently, invisibly out of date — the exact failure a context layer exists to prevent.

The fix is small: surface a last-synced timestamp everywhere the context gets read, by a human or an agent. Freshness you can't verify isn't freshness. It's a guess with better production values.