Fresh Beats Smart

A big platform just shipped a grounding service that answers a question most agent demos quietly dodge: how does your agent know anything that happened after training. The interesting part is not that it searches the web; plenty of things do that. It is how it hands the results back.

Instead of dumping whole documents into the context window and letting the model swim, it returns structured evidence objects and passage-level extracts, with citations, the source URL, and, crucially, when the page was last crawled and how stale it might be. That is a very different philosophy from “here are ten pages, good luck.” It treats grounding as a curated supply of evidence, not a firehose.

Three things worth stealing from that design, whatever tools you use.

Fresh beats smart. A slightly less clever model working from today’s data will beat a genius working from last year’s. Recency is not a nice-to-have bolted onto retrieval; for a lot of real tasks it is the whole value. If your agent cannot tell you how old its facts are, it cannot tell you whether to trust them.

Provenance is a feature, not a footnote. Citations and crawl timestamps are what let a human, or another agent, decide whether to believe an answer. Grounding without provenance is just confident guessing with extra steps.

And grounding is a token budget. Whole documents mean expensive inference downstream as the model re-reads the same noise over and over. Passages and structured evidence mean a higher signal-to-noise ratio and fewer tokens burned to reach the same answer. Retrieval quality is not just an accuracy lever; it is a cost lever, and the two pull in the same direction.

The headline race is still about which model is smartest. The quieter, more useful race is about which system feeds its model the freshest, best-cited, most compact evidence. Model IQ is converging. Grounding quality is not, and that is where a lot of the real product difference is going to live. Feed your agent better, not just bigger.