Microsoft Scales Agents via Sub-Agent Retrieval and Rubric Evals
July 14, 2026
Microsoft operates production agents in Foundry and Copilot by treating retrieval as a dedicated sub-agent. The system also utilizes rubric-based evaluations to maintain performance at scale.
HOW THIS AFFECTS YOU
●
builderYou can improve agent reliability by modularizing your retrieval steps into specialized sub-agent workflows.
●
researcherConsider implementing rubric-based evaluation frameworks to move beyond simple pass/fail metrics in production.