Foundry IQ went serverless. That's the part to notice.

Microsoft Foundry shipped a quiet but consequential update this week. The Foundry IQ post by Pablo Castro reads like a feature dump on first scan. Read the lines a second time and the actual story is more interesting: the knowledge layer for agents just got real production primitives, and the pricing model finally matches how agents actually consume retrieval.

The pieces worth flagging.

• Foundry IQ knowledge bases are GA, with a stable API, SLA, compliance certifications, and an MCP server that any MCP-compatible host can hit. That last bit matters; the “we wrote a RAG pipeline from scratch” pattern is now an unnecessary line item on most agent projects.
• Serverless retrieval is in public preview, scale-to-zero, billed in Compute Units. As Castro notes, agent workloads are “bursty and event-driven: an agent might execute hundreds of steps in seconds, then go idle for hours.” That shape was actively punished by the prior cluster-based pricing. Now it isn’t.
• A new multi-source knowledge base unifies Work IQ, Fabric IQ (including data agents and ontology), File Search, Azure SQL, and arbitrary MCP sources behind a single retrieval call. No custom integration code per source.
• Web IQ is available inside Foundry IQ, with sub-165ms latency and zero data retention. That last clause is the one regulated-industry SAs should circle.
• Security in preview: encryption controls, permissions sync, and sensitivity-label governance that carry through from source content into agent context. The Purview-aware pieces are the part that lets a real customer’s classified data flow into an agent without breaking the existing label scheme.

Read together, this is what the “system around the model” pitch from Build looks like in actual code form. Three practical reads for partners and customers.

• Stop hand-rolling RAG. A serverless multi-source retrieval endpoint with permissions sync and an MCP front door is now a platform primitive. The work that used to be “build a RAG pipeline” should move to “configure a knowledge base, point the agent at the MCP endpoint, and spend the saved time on evals.”
• Compute Units are the new line item. The honest budgeting question for an agent program now is “what’s our retrieval CU profile per session, and what is it under peak traffic.” If your TCO model still assumes flat-rate retrieval, refresh it.
• For SI and ISV partners, an MCP-aware Foundry IQ integration is a deliverable customers can recognize. A reference build that wires Foundry IQ Serverless + a multi-source knowledge base + sensitivity-label-preserving ingestion into a vertical workflow is the right shape of project for the back half of 2026. Customers will know what they bought.

The flashy AI work this week is going to be custom chips and Slack-native agents. The work that quietly compounds for customers is exactly this kind of knowledge-layer plumbing. Worth setting aside a half-day to run the cookbook.