Microsoft Build last week buried, somewhere between the seventeenth GitHub Copilot demo and the announcement of the RTX Spark dev box, a runtime announcement that is doing more work than the trade press gave it credit for. It is called Hosted Agents, it sits inside Foundry Agent Service, and it is in preview now with general availability targeted for early July.
The pitch in one sentence: every agent session gets its own sandbox, with dedicated compute, memory, and filesystem, scaled elastically, framework-agnostic, with the agent state persisted between turns. The internal Microsoft framing is that this is “the primitive for agents the way containers were for cloud-native apps,” which is the kind of phrase Microsoft says about everything but happens to be a useful framing here. Containers solved a real problem: process isolation without VM overhead, with reproducible environments. Agent sessions have an analogous mess. Right now most production agent deployments are some custom mashup of ephemeral compute, vector DB, queue, retry logic, and per-tenant state that the application team built from scratch and is afraid to touch. A managed runtime that ships all of that as a primitive is, structurally, the right product.
The model catalog underneath also got finalized at 11,000+ entries, which is a number that means almost nothing on its own but tells you what the strategy is: Foundry intends to be the place where you do not have to pick a lab. The shelf now stocks OpenAI’s GPT-5.5, Anthropic’s Claude Opus 4.8 and the rest of the Sonnet and Haiku 4.5 line, Google’s Gemini family, Microsoft’s own MAI models including the new MAI-Code-1 and MAI-Thinking-1, plus the entire Fireworks AI open-weight catalog, which is now generally available on Foundry as a single integrated surface with Azure data residency and enterprise governance. The selling proposition to large enterprise buyers is no longer “Azure has OpenAI.” It is “Azure has all of them, in one billing relationship, with the same SLAs.”
The bet underneath the bet is that the agent runtime is going to be a more important lock-in than the model itself. Models are interchangeable on a six-month cycle. The runtime your agents are written against is a multi-year decision. AWS announced something structurally similar with Bedrock AgentCore at re:Invent. Google has Vertex AI Agent Builder. The race is now visibly on for which hyperscaler owns the runtime primitive that production agents end up depending on, and Microsoft is the first to use the word “container” out loud about it.