What's the difference between Windows Agent Runtime, Apple Intelligence, and Android Gemini?

All three turn the operating system into an AI agent platform, but with different architectures and scope as of May 2026. Windows Agent Runtime (Microsoft, June 2026 Insider preview) is a new OS-level platform for registering, discovering, and invoking agents through standardized Windows APIs — initial preview is text-only on JSON/XML/PDF, vision agents deferred. Apple Intelligence on macOS, iOS, iPadOS routes through App Intents and Shortcuts — visual, deeply integrated, but agent capability is narrower and more conservative. Android Gemini (Gemini Intelligence) runs deeply integrated agents on-device with Gemini Nano plus cloud routing to Gemini 3.5 Flash — strongest screen-understanding capability, broadest hardware reach via Android. Pick by where your users live.

Which OS has the best AI agent platform in 2026?

It depends on the workload. For enterprise agents with deep policy controls, audit, and Microsoft 365 integration, Windows Agent Runtime plus Windows 365 for Agents is the most ambitious platform — though the June 2026 preview is still scoped narrowly. For Apple ecosystem consumers, Apple Intelligence has the cleanest cross-device sync and privacy posture but the most conservative agent capability. For broadest hardware reach and strongest screen-understanding, Android with Gemini Intelligence — Gemini Spark (always-on agent) plus on-device Gemini Nano spans the widest device range. As of May 2026 none have all three of vision, openness, and enterprise governance.

Can the same agent run on Windows, Mac, and Android?

Not directly. Windows Agent Runtime exposes Windows-only APIs. Apple Intelligence exposes Swift/Intent APIs for App Intents and Shortcuts. Android exposes the Gemini system services and on-device Gemini Nano APIs. The portable layer is the LLM and tool layer — most production cross-platform agents today are built on Anthropic, OpenAI, or Google APIs with frontend adapters per OS. Microsoft's Agent Framework SDK supports .NET and Python and is open source, which is the closest to a portable agent runtime, but it still targets Windows for OS-level integration. Cross-platform teams build the agent core once and per-OS thin adapters around it.

Does Windows Agent Runtime require Microsoft AI models?

No. The runtime is model-agnostic — agents can use Claude Opus 4.8, GPT-5.5, Gemini 3.5 Flash, local models via Windows Copilot Runtime, or any third-party model via API. Microsoft samples lean on Microsoft Foundry and Azure AI Foundry, but nothing in the platform layer forces a Microsoft model. Anthropic models are already integrated with Microsoft 365 Copilot with admin-toggle defaults for Excel and PowerPoint (Word coming summer 2026), and that pluggable model pattern extends to Windows Agent Runtime registrations.

Quick Answer

Windows Agent Runtime vs Apple Intelligence vs Android Gemini

Published: May 31, 2026

Windows Agent Runtime vs Apple Intelligence vs Android Gemini (May 2026)

Three OS-level AI agent platforms, three different bets on what “the operating system as agent platform” means.

Last verified: May 31, 2026.

TL;DR

Platform	Best for
Windows Agent Runtime (Microsoft)	Enterprise agents with policy controls, audit, Microsoft 365 integration. Most ambitious but earliest stage (June 2026 preview).
Apple Intelligence (Apple)	Consumers in the Apple ecosystem who want privacy, on-device, deep cross-device sync. Most conservative capability set.
Android Gemini Intelligence (Google)	Broadest hardware reach, strongest screen-understanding, on-device Gemini Nano plus cloud Spark. Best for general consumer AI assistant.

What each one actually is

Windows Agent Runtime

A new OS-level platform layer Microsoft is shipping in Insider preview in June 2026, announced at Microsoft Build 2026 (June 2–3, Fort Mason SF). The platform:

Lets apps register agents through Windows APIs.
Routes invocations between apps and agents.
Sandboxes execution with enterprise policy (Intune, Purview).
Ships an open-source SDK (Microsoft Agent Framework, .NET + Python).
Supports any model — Claude, GPT, Gemini, local.

Initial scope is intentionally narrow: text-only agents on JSON/XML/PDF. Vision-based agents (screen-pixel understanding) are deferred to a later release. Companion products: Windows 365 for Agents (cloud PCs for agent workloads), Windows Agent Store (distribution), Copilot agent mode (consumer surface).

Apple Intelligence

Apple’s OS-level intelligence layer, GA across macOS, iOS, iPadOS, and visionOS. The platform:

Routes most queries through on-device models (Apple’s Foundation Models, including a vision model and an LLM).
Escalates harder queries to Private Cloud Compute (Apple’s hardened cloud infrastructure with cryptographic attestation).
Optionally routes to ChatGPT, Claude, or Google via per-request extensions (user-confirmed).
Integrates with App Intents and Shortcuts as the agentic invocation surface.

Agent capability is conservative — App Intents let apps expose actions Siri and Apple Intelligence can call, but full autonomous multi-step agents are still rare. The May 2026 wave brings deeper App Intents for Mail, Calendar, Photos, plus stronger Visual Intelligence on macOS.

Android Gemini (Gemini Intelligence)

Google’s deeply-integrated AI layer across Android, GA. The platform:

Runs Gemini Nano on-device for low-latency local inference (summarization, reply suggestions, on-device privacy).
Routes to Gemini 3.5 Flash in the cloud for harder queries.
Powers Gemini Spark — Google’s always-on personal agent ($100/mo AI Ultra in US, May 30, 2026 launch).
Strong screen-sharing and screen-understanding capabilities.
Spans the broadest hardware footprint of any agent platform — phones, tablets, ChromeOS, Wear OS, Android Auto, Google TV, and Gemini-for-Home.

Side-by-side

Dimension	Windows Agent Runtime	Apple Intelligence	Android Gemini
Status (May 2026)	Insider preview June 2026	GA	GA
Hardware reach	Windows 11/12 desktops, Windows 365 cloud PCs	Mac, iPhone, iPad, Vision Pro	Android phones/tablets, ChromeOS, Wear, Auto, TV, Home
On-device models	Yes (Windows Copilot Runtime)	Yes (Apple Foundation Models)	Yes (Gemini Nano)
Cloud fallback	Azure AI Foundry, any provider	Private Cloud Compute + ChatGPT/Claude extension	Gemini 3.5 Flash, Spark
Screen/vision agent	Not in June preview	Limited (Visual Intelligence)	Strong (screen share to Gemini)
Cross-app agent registry	First-class (Agent Framework)	App Intents / Shortcuts	App Actions, intents
Enterprise governance	First-class (Intune, Purview)	MDM-managed	Workspace + Endpoint Mgmt
Pluggable models	Yes (Claude, GPT, Gemini, local)	Limited (per-request opt-in)	Google models only (mostly)
Always-on personal agent	No (not yet)	No	Yes (Gemini Spark)
Open SDK	Microsoft Agent Framework (open)	App Intents (Swift, closed runtime)	Limited

Where each one wins

Windows Agent Runtime wins when…

You’re building for enterprise Windows fleets with strict governance, audit, and policy.
You want a cross-app agent registry so agents can invoke each other without bespoke integrations.
You’re already on Microsoft 365 + Foundry + Azure and want the agent layer to align.
You can wait for the June 2026 preview and tolerate text-only-structured-data scope.

Apple Intelligence wins when…

Your users live in the Apple ecosystem and value cross-device continuity (Handoff, iCloud, Continuity Camera).
Privacy posture matters — Apple’s on-device + Private Cloud Compute is the strongest privacy story.
You’re shipping a consumer app where deep OS integration (Siri, Shortcuts) matters more than agent depth.
Your agent needs are conservative — augmentation rather than autonomous multi-step work.

Android Gemini wins when…

You want the broadest hardware reach — phones plus tablets plus ChromeOS plus Auto plus TV plus Home.
Screen understanding matters — Gemini can see what’s on screen and act on it natively.
You want an always-on personal agent today — Gemini Spark is GA (US AI Ultra) as of May 30, 2026.
You’re already invested in the Google Workspace stack.

The cross-platform problem

If you’re building an agent that should work on Windows, Mac, and Android — none of the three platforms makes this easy. The practical pattern in May 2026:

                Agent core (Anthropic / OpenAI / Google API + tools)
                                    │
                ┌───────────────────┼───────────────────┐
                │                   │                   │
        Windows adapter      Mac/iOS adapter     Android adapter
    (Agent Framework +     (App Intents +       (Gemini System
     Windows Agent          Shortcuts +          Services +
     Runtime invocation)    SiriKit)             Intents)

Build the agent core once on top of an LLM API (or Microsoft Agent Framework — it’s open source and the most portable of the SDKs). Build per-OS thin adapters for system-level integration. Don’t try to use a single OS-level platform across all three.

What changes at Build 2026

The big question is whether Microsoft’s Windows Agent Runtime + Windows 365 for Agents combination becomes the dominant enterprise agent platform. The argument for: deeper governance than Apple or Google, ISV ecosystem motivation via Agent Store, and integration with the world’s deployed Microsoft 365 base. The argument against: shipping text-only-structured-data first means vision-heavy agentic computing (which is where consumer wow factor lives) is deferred.

Watch Nadella’s June 2 keynote for the framing, then the Build 2026 sessions on Agent Framework, Foundry, and Windows 365 for Agents for the technical depth.

Quick decision tree

Where do your users live?
├── Windows enterprise → Windows Agent Runtime + Foundry + 365 for Agents
├── Apple consumer → Apple Intelligence + App Intents
├── Android consumer → Gemini Intelligence (+ Spark if always-on agent)
└── Cross-platform → Anthropic/OpenAI/Google API + per-OS adapters

Verdict

No single OS-level agent platform wins in May 2026. Microsoft is most ambitious on enterprise governance and cross-app registry but is earliest stage. Apple is strongest on privacy and on-device but most conservative on agent capability. Google has the broadest reach and strongest screen-understanding but is least open. For pure platform bets, watch Microsoft Build 2026 — Windows Agent Runtime is the most strategically important agent-platform announcement of the year. For shipping cross-platform agents today, the right answer remains: build on an LLM API, integrate per-OS where it matters, and don’t lock in to one platform’s runtime yet.

Sources: Microsoft Build 2026 conference page, Windows News Agent Runtime preview coverage, Apple Intelligence developer docs (developer.apple.com/apple-intelligence), Google I/O 2026 Gemini Intelligence keynote, PCMag Build 2026 preview, Yaabot Gemini Intelligence Android guide, Help Net Security Windows 365 for Agents coverage (verified May 31, 2026).