Microsoft Foundry's July and August 2026 update roundup brings Hosted Agents and Toolboxes to general availability, adds Claude capabilities to Azure-hosted deployments, and expands Model Router with new regions and models. The changes were announced in a blog post by Nick Brady on September 9, 2026, covering updates from the two-month period.
Hosted Agents, which let developers build agents with their preferred framework and run them in Foundry's managed runtime, reached GA on July 9. Toolboxes, which provide a single MCP-compatible endpoint while Foundry manages tool authentication and credentials outside agent code, also became generally available. Voice Live integration for hosted agents is now GA as well.
Confirmed
According to the Foundry blog, the August 17 announcement brought structured outputs, Web search, Web fetch, MCP connector, and Tool search to Claude models hosted on Azure. The MCP connector uses the beta API. These capabilities turn a model endpoint into a production agent platform, according to a separate post by Haoran Cheng.
Model Router expanded in August with additional regions, a refreshed routing pool including GPT-5.6 variants and Claude Opus 4.8, and agentic routing for eligible Anthropic and open-source models. Foundry Local on Azure Local added preview evaluation and multi-GPU inference capabilities. Extension 2607 brought model evaluation, vLLM model parallelism, and improved automatic GPU inference tuning.
Foundry DevPack shipped preview installers in the August 0.1.3 release for Windows, macOS, and Linux on x64 and Arm64. By the end of August, Python and JS/TS SDKs reached stable version 2.5.0, Java 2.4.0, while.NET's 3.0.0 line remained in preview.
Unknown
Hosted-agent management and runtime support differ by language, and the blog notes that developers should check stable-client changes and runtime requirements before migrating existing agents. The post does not provide independent benchmarks or third-party validation for the new capabilities. Pricing details for the new Claude tools and Model Router updates are not specified in the roundup.
Tool Search, which reduces token use by retrieving only relevant tools at runtime, is in public preview. Agent Skills and tool search are also in public preview. The blog does not state when these features will reach general availability.
Our take
Microsoft's push to make Foundry a managed runtime for agentic workloads is clear with the GA of Hosted Agents and Toolboxes, but the strategic value may lie in Tool Search. As tool catalogs expand, the overhead of sending every schema on every turn creates a significant token and context bottleneck. Runtime retrieval of relevant tools addresses this scaling challenge directly, though the absence of independent performance data means the actual impact on latency and cost remains unproven.
