A bulk reindex made updated_since useless for release-date queries, and high IDs turned out not to mean recent releases.
An AI agent asked "which products in this category came out recently?" produced confident, wrong answers. Two shortcuts were to blame.
Shortcut 1: updated_since. It means "when this row last changed," not "when this item came out." Worse, a bulk reindex had touched nearly every row — so "updated recently" meant almost nothing. An item from the 1990s and one released last week carried nearly identical timestamps.
Shortcut 2: highest IDs. Auto-increment IDs record when a row was inserted, not when the thing happened. Catalogs are backfilled constantly; a high ID can be an old item catalogued last month.
The fixes
- Answer release questions from actual release-date fields — filter by category, constrain on date, then fetch details in one batch.
- Fix the tool descriptions: say what each field means and what it must not be used for. The agent chose
updated_sincebecause the description let it.
The lesson: agents are good at plausible plans and bad at knowing which fields are trustworthy. Field semantics are part of your API contract — document them like one.