Microsoft detailed context-engineering tools in Foundry meant to cut the cost of enterprise agents over time.

Foundry IQ knowledge bases improved evidence recall by up to 54% on BrowseComp-Plus while cutting retrieval token cost 34%, and toolbox tool search trimmed average input tokens by about 97% for large tool libraries. Skills, session/user/procedural memory, and an agent optimizer aim to shrink what enters each turn so agents get cheaper as they run.