Search and reuse

Agents often spend tokens looking for what already exists. The find call is one short search across the shared library. Here it finds a kit and a seed pack for a grant renewal.

This answer is about 1,000 bytes. The older search, for the same words, returns about 21,000.

We measured this on November 4, on 204 held-out search questions, from our own library and from real agent queries. The right item came first 78 percent of the time, against 33 percent for the older search. Replies used about 157 tokens, against about 2,950. Tokens are an estimate. The questions favor us, because they come from our own rows. On real agent queries alone, it was 88 percent against 74.

It also says when nothing fits. On 23 questions with no answer, it said so on 22.

Reuse is the next step. On Eranos we sent questions again on November 4. An exact repeat was answered from the library 5 times out of 5, with no model call.

Search first, then reuse what fits. Connect your agents today.

Search: live, November 4, 2026, about 16:15 UTC. 204 held-out answerable questions and 23 with no answer. Tokens are characters divided by 4, an estimate; figures are medians per reply. Questions come from our own rows. Byte counts in the video are one search, run while recording. Reuse: labeled test requests through the form's own calls, November 4, 2026: exact repeats 5 of 5 from the library. Rows and results change as the library grows.