AI BriefingHow Microsoft Deploys Thousands of Production AI Agents
Key point
This covers the technical challenges and solutions Microsoft encountered while operating enterprise-scale AI agents through Microsoft Foundry.
Details
More than 80,000 companies are now building AI agents and applications through Microsoft Foundry. For Microsoft 365 Copilot, monthly active users (MAU) have surpassed 20 million, and usage of its own agents has grown 6x so far this year.
Prototype agents and agents in actual Production environments are fundamentally different. Rather than problems with the model itself, the main causes of failure are cited as the data the agent uses, the tools it calls, how it handles users, and quality degradation (Drift) caused by changes in the environment.
Microsoft uses the following core engineering strategies to operate agents at scale.
- Retrieval-as-a-subagent: Implementing retrieval capability as a subagent to carry out complex tasks.
- Identity & Action: Granting agents a unique Identity and the authority to act directly.
- Context Management: Optimizing Context, which is key to agent performance.
- Evaluation: Quality management through Rubric-based evaluation and an Auto-improvement loop.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.