Journal

An agent that writes analytics reports will eventually invent a statistic. Ours announced a 26.7% reply rate no query produced. This is how Spot's report generation was rebuilt so a report cannot state a number it didn't compute — and the blind, adversarial scoreboard that proves it.

A capable model with ten tools burns through its step budget before it can answer anything. So Stir routes every query through a tiny classifier first, scopes the main model to a four-tool subset per intent, and short-circuits one-word queries with zero LLM calls.