What changed
The adoptable agent development this cycle is a runtime, not a model. While GPT-5.6 sits behind a roughly 20-partner limited preview and Gemini 3.5 Pro stays limited-preview, Microsoft Foundry’s Hosted Agents (Foundry Agent Service) is a managed, framework-agnostic runtime that runs your own agent code, packaged from any framework, on Microsoft-managed infrastructure 12. It is callable today in public preview across 20 Azure regions, and Microsoft targets General Availability by end of June 2026 1; the Microsoft Learn concept page, updated 25-06-2026, still labels it “currently in preview” 2, so the honest framing is preview-now, not GA-shipped. The core fact is dual-sourced: the docs and an independent InfoQ write-up both describe managed sandboxed sessions with persisted state and filesystem access, multiple-framework support, and a stateful Responses API plus a lighter Invocations protocol 23.
The runtime layer
In my read, the agent stack is bifurcating into three layers - model, framework, and runtime - and the runtime (where the agent actually executes: isolation, identity, persisted state, scaling) is becoming a product you rent rather than hand-build. The engineering substance backs that up. Each session runs in its own VM- or hypervisor-isolated sandbox with a persistent $HOME filesystem, enabling scale-to-zero with stateful resume 12. Every deployed agent is auto-assigned its own Microsoft Entra ID (agent identity) and a dedicated endpoint at deploy time 12. And you deploy from source, no hand-built container required 2:
# Deploy agent source directly (no container packaging required)
azd deploy
# Each deployed agent is auto-assigned a Microsoft Entra ID (agent identity)
# and a dedicated endpoint at deploy time.
# Protocols exposed: Responses (OpenAI-compatible /responses), Invocations (JSON).
The operational envelope
These are Microsoft-documented numbers, single-sourced to the Learn docs and not independently corroborated; treat them as a starting point to verify against your own load.
| Limit or knob | Documented value |
|---|---|
| Idle timeout | 15 minutes 2 |
| Session lifetime | deleted after 30 days inactivity 2 |
| Default quota | 50 concurrent active sessions per subscription per region 2 |
| Sandbox sizes | 0.5/1/2 vCPU (1/2/4 GiB) 2 |
| Per-session disk | up to 20 GiB (~20% reserved for system) 2 |
Impact on your team
If you hand-build agent hosting today (isolation, identity, state, scaling), this is the layer it replaces, and unlike this cycle’s gated model drops it is callable now, so it is worth a spike. The catch is where the lock-in goes. The agent code stays genuinely framework-agnostic (Agent Framework, LangGraph, Semantic Kernel, or custom Python/C#), but the runtime and identity are hard-wired to Azure and Microsoft Entra, so the lock-in moves from your code to your runtime and IAM. That is the real architecture decision, not the demo. Two concrete watch-items before any rollout: it is preview, with GA only targeted for end of June 2026, so do not put a production SLA on it yet; and because billing scales per session, you right-size the sandboxes first, not after.