Fewer Searches Before the Right Tool ​

September 4, 2026 · InfraScout Team

InfraScout exposes 221 tools and does not hand all of them to the model on every turn. They load on demand, which means the model searches the catalog for what it needs.

That works when the search returns the right tool. When it does not, the model tries again with a different pattern, and every retry re-reads the whole conversation — which is the part you are charged for. A turn that should have cost one search can cost eight.

Searches were being cut short with nothing to say so ​

A catalog this size returns more matches than a single response should carry, so results are capped. The cap was landing well short of what a broad question needed, and nothing in the result signaled it.

A search for Intune device tools would come back with 5 of the 21 that existed at the time, the target not among them, and no way for the model to tell that apart from a search that genuinely found nothing relevant. So it searched again, and again.

Searches now state how many results they want. Across real assessment turns, the number of searches per turn dropped from roughly 21 to 11 — half the re-reading of your conversation on every turn that has to look a tool up.

The default model could not search at all ​

If you have not configured your own AI connection, your conversations run on Claude Haiku 4.5. That model had been excluded from tool search entirely.

So the default could not search. Every new conversation on it fell back to choosing tool groups up front from your first message alone, which costs an extra step before the conversation starts and can leave out a capability the conversation turns out to need.

It searches now, like every other model, and those conversations start the same way the rest do.

What you should notice ​

Fewer stalls at the start of a turn, and less of the AI visibly casting around for a tool you know exists. If you run without a configured AI connection, new conversations should also start more directly.

Nothing to configure. Both changes are live for every tenant.

Questions or feedback? Reach us at info@infrascout.cloud.