From Good to Great: We Put Agent Optimizer to the Test in Microsoft Foundry

Authors: Abi Komma, Charles Kim, Luffy Chen, Morteza Ziyadi Why hand-tuning agents does not scale An agent’s behavior comes mostly from its natural-language configuration: its instructions, skills, and tool descriptions. Most teams tune that configuration by hand. You write the instruction , run some cases, read the failures, reword , and try again. The Foundry documentation ( link ) calls this the manual trial-and-error loop. It is slow, subjective, and hard to scale across many agents. The loop is also brittle. Unlike a code bug, a quality failure has no stack trace: the cause can sit in the