On June 3 through 5, Microsoft pushed about 40 Azure updates. Microsoft Foundry IQ went GA, Microsoft Discovery went GA, HorizonDB and DocumentDB kept extending, Azure Monitor kept closing the OpenTelemetry gap, and a cluster of agent-side updates landed that line up directly with the Agent Framework posts from June 1 and June 2. The Agent Framework follow-on is the most useful framing for the drop, so it goes first.

Follow-on to the Agent Framework posts

If you followed the Agent Framework posts last week, six of these updates are direct follow-ons. Part 1 walked from a single ChatClientAgent to one wired up to a remote MCP server. Part 2 ran three of them in a debate and gated a Jenkins deploy on the verdict. This week Microsoft shipped first-class versions of most of the moving parts those posts hand-rolled.

Agent-to-agent (A2A) for Prompt and Hosted agents in Foundry (preview)

The Part 2 debate was hand-rolled: three ChatClientAgents sharing a transcript across three rounds, with a judge persona constrained by a strict-format prompt. A2A is Microsoft formalizing exactly that pattern as a first-class Foundry primitive for prompt agents and hosted agents. Worth watching whether the preview converges on the same multi-round, single-transcript shape the debate post used, or on a different topology.

Rubric evaluator and Intelligent Trace Sampling evaluations (preview)

The judge persona in Part 2 was a hand-built rubric (VERDICT: APPROVE / BLOCK / NEEDS_HUMAN plus a confidence score). The rubric evaluator is the productized version of that pattern. Intelligent trace sampling is the answer to the question that breaks most agent-eval pipelines: how do you score the runs without paying to evaluate every single one. It picks representative traces instead of evaluating wall-to-wall.

Foundry for VS Code (GA, June Build 2026 refresh), code-first observability for Foundry Agents in VS Code (preview), and observability developer experience in Azure Developer CLI (preview)

Both Agent Framework posts were code-first Python against an agent-framework==1.0 install, run from a terminal. This week is Microsoft shipping the IDE and CLI dev loop for that exact workflow. The VS Code refresh is GA; the in-editor agent observability and the azd observability experience are both in preview. Pair them and the local agent dev loop in VS Code starts to look like the local API loop did five years ago.

Tool search in Foundry toolboxes (preview)

Part 1, example 3 connected to the Microsoft Learn remote MCP server as a tool source. Tool search is the catalog UX for that pattern once a Foundry toolbox holds more than one MCP server. Useful the moment a team has more than two or three.

Microsoft Foundry IQ (GA) and two new knowledge sources (preview)

Foundry IQ is the managed knowledge layer for grounding agents in enterprise data: connect SharePoint, OneLake, Azure Blob, and other sources, and Foundry IQ handles the retrieval pipeline that previously had to be rebuilt for each project. The MCP-server-as-tool-source pattern from Part 1 covers half of grounding; Foundry IQ packages the retrieval half behind one managed surface. The two new sources land alongside the GA: Azure SQL Database becomes a first-class knowledge source, and a Microsoft Fabric Ontology can be queried as a federated source. The Fabric Ontology one is the more interesting half. Agents query the semantic layer Fabric customers already curate, instead of a parallel definition built only for retrieval.

User feedback logging in Microsoft Foundry (preview)

Part 2 writes a PR comment and exits 0, 1, or 2. The Jenkins gate decides the next step; the eventual human verdict on whether the gate was right never makes it back into the agent. Feedback logging is the Foundry-native equivalent for capturing that verdict so the next evaluator run has a ground-truth signal to train against.

The pattern across all six is the one the agent space has been on for a year. A community or hand-rolled approach gets validated, then absorbed into a managed surface.

Microsoft Discovery (GA)

Microsoft Discovery is generally available as an enterprise platform for building and governing agentic AI workflows for R&D organizations across scientific and engineering disciplines. This was previewed earlier in the year. The GA is the signal that the agentic-workflow surface inside Microsoft is no longer just the Foundry Agent Service.

Foundry and AI Search platform updates

Four more Foundry and AI Search items that are not Agent Framework follow-ons but are worth knowing.

Private connectivity for AI Search and Foundry Knowledge Bases (GA)

Ingestion, enrichment, retrieval, and agent traffic between AI Search and Foundry Knowledge Bases can now stay on Shared Private Link or Network Security Perimeter end-to-end. Together with the Purview integration that went GA in the June 2 drop, the retrieval layer is closing the same governance and networking gaps the data-plane services closed years ago.

APIM support for Foundry Models in Azure AI Search (preview)

Azure API Management can now front all Foundry model integrations used by Azure AI Search. The reason this matters for platform teams: it puts a single throttling, key-vault, and observability surface between AI Search and the underlying model deployments, instead of each search workload calling models directly.

Content Understanding chunking and image verbalization in AI Search (preview)

The Content Understanding pipeline can now chunk and verbalize images as part of AI Search indexing. The output is searchable text derived from images, which lets a single retrieval query span text and visual content.

Domain filter in the Foundry model catalog (preview)

The Foundry model catalog adds a domain filter that narrows the 1,900-plus models to the ones trained for a specific industry or use case, with filters for domains like robotics and biomedical sciences. A small UX change, but the catalog crossed the size where browse-by-name stopped scaling a while ago. This is the model-catalog equivalent of the tool search update above – the same problem (too many things in the catalog) solved one layer down.

Databases

HorizonDB AI pipelines (preview)

HorizonDB, the Postgres-compatible database introduced in the June 2 drop, now lets you describe an AI ingestion workflow (chunking, embedding, extraction, generation, ranking) declaratively in SQL and run it as a fault-tolerant pipeline inside the engine. Same play as the rest of HorizonDB: keep the RAG pipeline on the operational database instead of stitching together a separate service per stage.

DocumentDB advanced full-text search (preview)

Advanced full-text search lands in DocumentDB, alongside the instant free-tier clusters that shipped on June 2. HorizonDB and DocumentDB are clearly the two databases Microsoft wants to push for new workloads.

Postgres Flexible Server DuckDB extension (GA)

The DuckDB extension is now GA in Azure Database for PostgreSQL Flexible Server. DuckDB-in-Postgres turns the operational instance into a competent analytics endpoint for parquet and CSV in blob storage without moving the data. For the small-to-medium analytics that do not justify a Fabric or Synapse footprint, this is the simplest viable answer.

Azure Monitor

Three Monitor updates landed together and all went GA.

OTLP ingestion is GA: send OpenTelemetry Protocol signals straight from instrumented applications and platforms to Azure Monitor with no Application Insights SDK in between. Dynamic thresholds for log search alerts went GA, so the platform calculates the threshold instead of asking the operator to guess. And Azure Monitor Service Level Indicators reached GA. Combined with the simple log alerts and OpenTelemetry metrics that shipped on June 2, Azure Monitor is methodically closing the gap with the open observability stack.

A day later, on June 5, Metrics Usage Insights added an Ingestion Volume Change dashboard in preview, for comparing ingestion volume over time and spotting spikes or drops in time-series counts and event ingestion rates. It is a cost-and-noise control surface more than an observability one – the dashboard you open when the Monitor bill jumps and you need to know which stream moved.

Confidential computing

Three updates on the confidential side. Confidential Clean Rooms gets a preview of multiparty analytics, a managed service for partners to jointly analyze privacy-sensitive datasets with Apache Spark without exposing the underlying data. Confidential live migration for Intel TDX VMs is in development, which is the last big operational gap separating confidential VMs from regular ones. And Azure Confidential Ledger gains a GA backup tool for audit retention of ledger files.

GitHub Copilot modernization agent (GA)

The GitHub Copilot modernization agent is GA. It coordinates application assessments and upgrades across a whole portfolio, not just a single repo. For the Java-on-old-Spring or .NET-Framework migration backlog most enterprises still carry, this is the first serious estate-wide automation Microsoft has shipped.

Migration

Azure Files assessments worldwide in Azure Migrate (GA)

Azure Migrate now discovers and assesses SMB and NFS file shares hosted on Windows and Linux servers, worldwide. File shares were the awkward gap in most migrate-to-Azure-Files plans: you could assess the servers but had to size and plan the share targets by hand. This closes that gap and gives a data-driven view of the file-share estate. Read it alongside the Copilot modernization agent above – one is the estate-level story for application code, this is the estate-level story for file data.

App Service Flex Consumption: rolling updates (GA)

Rolling updates are GA in the Flex Consumption plan. Instead of restarting all instances during a deploy, the platform rolls them. Zero-downtime deploys on Flex no longer need a slot-swap or external front door.

Compute

Lasv5 and Laosv5 storage-optimized VMs (preview)

Storage-optimized VM series based on the 5th-generation AMD EPYC (Turin). Lasv5 targets high disk capacity, throughput, and I/O. Laosv5 targets the same shape with a different storage profile.

Azure Infrastructure Resiliency Manager (preview)

A new preview service for orchestrating resiliency testing and recovery across an Azure estate. It sits in the same conversation as Chaos Studio but framed for the resiliency-program owner rather than the SRE writing fault-injection experiments.

Guest RDMA on Azure Boost (private preview)

Landing on June 5, Guest RDMA is in preview on Azure Boost, starting in UK South, bringing high-throughput, ultra-low-latency RDMA networking directly into guest VMs within a region. Offloaded to Azure Boost, this is the kind of networking that used to require specialized HPC SKUs showing up as a general guest-VM capability – the part that matters for tightly-coupled HPC and AI-training traffic that is sensitive to latency between nodes.

Speech, voice, and language

MAI-Voice-2 is in preview in Foundry. Custom Voice portal and self-serve custom photo avatar creation both went GA. Voice Live API also picks up avatar voice sync with custom voices in preview, pairing a branded or persona-specific text-to-speech voice with a real-time avatar – the piece that ties the custom-voice and custom-avatar tracks together. On the language side, Text Analytics for Health NextGen Playground is GA, and the Conversational and Text PII NextGen playgrounds shipped updates. The pattern: the NextGen playgrounds are now the default front door for the language services.

Region

Azure Red Hat OpenShift is GA in Belgium Central.