Microsoft’s Build 2026 message was less about a single new model and more about making agents run as managed software. On June 2 at Fort Mason in San Francisco, the company described Foundry, the former Azure AI platform, as a place where enterprises can host agent runtime instead of assembling isolation, state and storage themselves.
The most concrete piece is Hosted Agents, expected to become generally available in early July. Microsoft describes it as managed sandboxed sessions with state, durable file-system access and framework flexibility, which means developers can focus on what the agent does rather than rebuilding its execution environment.
Microsoft also filled in memory and voice. Procedural memory, user memory and session memory are in public preview, with procedural memory said to add 7 to 14 percentage points of absolute task success at near-baseline cost. Voice Live is generally available, combining speech recognition, text-to-speech, turn detection, interruption handling and virtual avatars in one API, while Foundry IQ and SLA-backed knowledge retrieval cover the retrieval layer.
The strategic layer is governance. Microsoft open-sourced ASSERT, introduced the Agent Control Specification, and added a rubric evaluator plus Guided Guardrail Setup. With Fireworks AI also arriving in Foundry and MAI models entering preview, Microsoft is positioning the platform as the place where agents are hosted, measured, guarded and billed, regardless of which model a company chooses.
Sources: CocoLoop, Microsoft Foundry Blog