Siri AI Goes Hybrid, MCP Goes Stateless

<p>Apple debuted <strong>Siri AI</strong> at WWDC (June 8) — a rebuilt assistant pairing on-device Apple models with a custom Google Gemini backend, making iPhone a first-class agent runtime. Meanwhile, the <strong>MCP Release Candidate</strong> drops session state entirely, and the Vertex AI SDK enters read-only mode on June 24 with a migration deadline ...

Highlights June 10

  • Apple unveiled Siri AI at WWDC 2026 (June 8) — a ground-up rebuild routing cloud intelligence through a custom Google Gemini model, making iPhone 16+ and iPhone 15 Pro the new mobile agent surface for the industry to target (CNBC)
  • The MCP Release Candidate (published May 21) removes the initialize handshake and session header entirely, making the protocol stateless at the transport layer — a breaking architectural shift SDK maintainers have ten weeks to absorb before the July 28 finalization (MCP Blog)
  • Anthropic Claude Managed Agents added Auto mode for injection screening, Worktrees for isolated branch execution, and Routines that trigger via scheduled intervals, webhooks, or API endpoints — framing agent infrastructure as the "new bottleneck, not intelligence" (InfoQ)
  • Google's Vertex AI SDK enters read-only mode on June 24 — teams still importing from the old SDK have two weeks to migrate to the renamed Gemini Enterprise Agent Platform (Google Cloud Docs)

Key Signals

  1. Signal 1

    Apple's Siri AI is a hybrid on-device/cloud agent

    June 8, WWDC 2026

    Apple shipped a fully rebuilt Siri — now branded Siri AI — that runs privacy-sensitive tasks on Apple's own foundation models and hands complex reasoning to a custom Google Gemini model in Apple's data centers. The assistant supports a dedicated Siri app with conversation history and cross-app actions, with a waitlist rollout limited to iPhone 15 Pro and the full iPhone 16 and 17 lines. For platform builders the signal is strategic: Google just landed the biggest cloud-AI distribution deal of the year, and mobile hardware has become a first-class agent runtime. (CNBC, Digitimes)

  2. Signal 2

    MCP is going stateless — the biggest spec change since launch

    RC published May 21, final spec July 28

    The next MCP specification removes sticky sessions at the protocol layer, adds Streamable HTTP headers so gateways can route without body inspection, and codifies ttlMs/cacheScope on list and read results for safe client-side caching. Two official Extensions ship alongside: Tasks (long-running async operations) and MCP Apps (server-rendered UIs). Roots, Sampling, and Logging are deprecated with a 12-month removal runway. Teams building MCP servers or client SDKs should audit session-state assumptions now. (MCP Blog)

  3. Signal 3

    Anthropic expands Managed Agents with orchestration primitives

    Code with Claude event, May 2026

    Alongside the April 8 public beta, Anthropic's Code with Claude event detailed additions integrating with Managed Agents: Auto mode screens for destructive actions and prompt injection before any tool call runs; Worktrees give Claude isolated git branches per task; Routines schedule or webhook-trigger agent execution without a persistent shell session. Dario Amodei framed the next inflection point as "teams of agents working at the level of organizations" — Notion, Rakuten, and Sentry are already in production. (InfoQ, InfoWorld)

Why It Matters / What To Watch

  1. Mobile is now an agent deployment target you can't ignore

    • Siri AI's hybrid architecture (on-device Apple models + Gemini cloud) will pressure competing platforms — watch Android Gemini and Windows Copilot Runtime to accelerate on-device agent APIs in response. (CNBC)
    • Whether Apple opens Siri AI to third-party agent invocation via App Intents is the next key disclosure — that would make iOS a distribution channel for enterprise agents. (HotHardware)

  2. MCP server authors have until July 28 to audit session state

    • The stateless core removes Mcp-Session-Id from the protocol — any server or proxy that tracks per-session routing state will need re-architecture before the spec finalizes. (MCP Blog)
    • Authorization hardening mandates iss parameter validation and tighter OAuth 2.0/OIDC alignment — check MCP server auth flows now if you run user-facing MCP deployments. (MCP Blog)

  3. Vertex AI SDK sunset is June 24 — two weeks away

    • All Vertex AI SDK modules stop receiving updates on June 24, 2026. The migration target is the Gemini Enterprise Agent Platform SDK, where Agent Runtime now supports up to 7-day long-running operations. (Google Cloud Docs)

Quick Links