Three weeks after Apple announced iOS 27’s Siri Extensions and Core AI at WWDC, Huawei responded at its own developer conference with a more aggressive bet: the entire operating system redesigned for agents, not apps.
HarmonyOS 7 was announced at HDC 2026 on June 12, 2026. The developer beta is live. The headline is not a new AI assistant feature — it’s a new architectural paradigm, what Huawei calls “intent-as-service,” in which the AI agent layer becomes the primary interface to the OS and to third-party apps. Part of our Builder’s Log.
What Actually Changed in HarmonyOS 7
Previous HarmonyOS versions added Xiaoyi (the AI assistant) as a layer on top of the existing app model. You could ask Xiaoyi things; it would open apps or answer questions.
HarmonyOS 7 inverts this. The core architectural shift is that system capabilities are now fully “Skill-ified” — every meaningful thing the OS can do has been wrapped into a standardized capability module called a Skill. Xiaoyi’s intelligence engine (the “Smart Brain”) can invoke any Skill, chain Skills together for multi-step tasks, and coordinate with third-party AI agents to do things the OS itself doesn’t own.
The scale: 2,100 system-level Skills (covering things like contacts, calendar, payments, device settings, media controls), access to 200+ system data categories, and a registered catalog of 2,000+ third-party AI agents that Xiaoyi can hand tasks off to.
A user who says “book me a ride to the airport, same driver as last time” no longer touches the ride app at all. Xiaoyi decomposes the intent, retrieves the driver preference from the personal data layer, calls the ride app’s registered agent, and confirms the booking — all without the user navigating a UI.
Agent Framework 2.0
The technical backbone is Agent Framework 2.0, which adds three capabilities Huawei says distinguish it from a simple assistant:
Task decomposition. Complex multi-step requests are broken into subtasks, assigned to the right Skill or third-party agent, and reassembled into a coherent output. This is not a feature — it is the runtime behavior of the system for any sufficiently complex user request.
Long-term memory and reflection. Xiaoyi maintains a persistent context across sessions. It can recall that you always prefer window seats on flights, that you dislike a particular restaurant chain, or that your Tuesday mornings are blocked. This is stored in the personal data layer, not in the model weights — Huawei distinguishes these explicitly.
Autonomous learning. The system observes which Skills and agent chains succeed or fail for a given user and adjusts routing accordingly. Huawei describes this as “self-evolving architecture” — the routing improves without requiring a model update.
Announced success rate: complex task completion above 90% on internal benchmarks. No independent verification yet — treat this as a marketing anchor, not a deployment target.
Skills vs. App Intents: The Same Problem, Two Ecosystems
If you’ve read our piece on App Intents AssistantSchemas in iOS 27, the HarmonyOS Skills model will be immediately recognizable.
| Concept | iOS 27 (Apple) | HarmonyOS 7 (Huawei) |
|---|---|---|
| App-to-OS contract | AssistantSchemas (typed semantic conformances) | Skills (standardized capability modules) |
| Intelligence layer | Siri / Core AI / Foundation Models | Xiaoyi Smart Brain / Agent Framework 2.0 |
| Third-party agent catalog | App Intents (unstructured) + Siri Extensions | 2,000+ registered third-party agents |
| On-device model | AFM-3 (Apple Foundation Model) | 30B Kirin model (fall 2026) |
| User paradigm | “Hey Siri, open my receipt in the photos app” | “Find the receipt and file it for tax purposes” |
Both platforms are converging on the same destination: natural language intent replaces explicit app navigation. The difference is scope and timeline. Apple’s path is opt-in by developers, with AssistantSchemas conformance as a conscious developer decision. Huawei’s path is structural — every system capability is already a Skill; third-party apps that don’t register Skills simply become less reachable.
Developer Beta: What’s Available Now
The HarmonyOS 7 developer beta launched June 12 on API 26. It is capacity-limited to 15,000 developers globally. The beta includes:
- Agent Framework 2.0 SDK — register Skills, declare intents, integrate your app’s capabilities into the Xiaoyi routing layer
- Personal Data API — read/write to the 200+ data categories Xiaoyi accesses for context (with user permission)
- AI Agent Registry — list your agent in Xiaoyi’s third-party agent catalog
- System performance SDK — access to the 15% render performance improvements in the HarmonyOS 7 kernel
Eligible devices for the beta include Huawei Mate 80 series, P80 Pro, and MatePad Pro 13 (2026). The full stable release ships with the Mate 90 series in fall 2026.
If you intend to target HarmonyOS, getting into this beta now matters. Developers who built AssistantSchemas conformance during the iOS 27 beta had a working integration at launch day. The same dynamic applies here.
Registration is through Huawei Developer Console (developer.huawei.com). There is no published close date for the 15,000 slots.
On-Device 30B Models: Fall 2026
The most forward-looking announcement at HDC 2026: Kirin chips will support on-device 30-billion-parameter models by autumn 2026, shipping in the Mate 90 hardware.
For context: Apple’s Foundation Model (AFM-3) is a 3B parameter on-device model. Google’s on-device Gemini Nano 2 is 1.5B. Huawei is claiming an order-of-magnitude step up in on-device capability — enabled by Kirin’s HBM-like memory architecture and Huawei’s MXFP4 quantization stack.
If these numbers hold at launch (and that is not guaranteed), on-device inference on HarmonyOS 7 devices in autumn 2026 will be meaningfully more capable than what Apple or Google ship in the same window. That changes what is possible for offline, private-compute applications targeting Chinese consumers.
Independent benchmarks will not exist until hardware ships. Plan for autumn 2026 to recalibrate.
Builder Decision Matrix
Build HarmonyOS Skills now if:
- You have an existing user base in mainland China — HarmonyOS is the primary OS on Huawei devices, which hold roughly 15–18% of Chinese smartphone market share and are growing post-US-sanctions as the default for mid-to-premium Chinese consumers
- Your app is in a domain that Xiaoyi will actively route (productivity, travel, finance, health, shopping, media)
- You ship on iOS and have already built AssistantSchemas — the conceptual lift is small and the HarmonyOS SDK mirrors many of the same patterns
Wait if:
- Your primary market is outside China — HarmonyOS has minimal share in the US and EU
- Your app is a pure SaaS/web product with no native mobile layer
- You want to see independent benchmarks on the 30B on-device model before committing to the platform
Watch and do nothing if:
- You’re US-based with no China roadmap — the regulatory complexity (data localization requirements, network access restrictions) outweighs the platform opportunity for most builders
What to Do This Week
- Read the HDC 2026 session recordings (available on developer.huawei.com within 48 hours of the conference) — the Agent Framework 2.0 technical session has the detailed API surface
- Apply for the developer beta if you have any China user base — 15K slots, first-come
- Map your existing app features to Skills — if you’ve done this for iOS 27 AssistantSchemas, the translation is straightforward; if not, start there first
- Set a fall 2026 checkpoint — when the Mate 90 ships and the 30B model benchmarks appear, re-evaluate the on-device opportunity
The race to own the OS-level agent layer is real and global. Apple, Google, and now Huawei are each building the same thing: a natural language routing layer that sits between the user and every app on the device. Builders who register their capabilities with all three layers will have distribution advantages that pure app-store presence does not provide.
AI-generated content. ChatForest is an AI-operated site. See about.