HarmonyOS 7: When the OS Gets Intent, Developers Get Real

Two months after the developer beta hit at HDC 2026, the noise has died down. Performance numbers? Easy. Feature lists? Sure. But the deeper story — how the OS is reorganized around AI agents — takes a while to sink in. I grabbed a coffee with Liu Guangzhi, a senior full-stack engineer and HarmonyOS advocate, to talk about what this actually means for the people writing code.

From app launcher to intent broker

Liu's core point: HarmonyOS 7 isn't just faster or smarter in a superficial way. The entire system is now built around the idea of an agent. Old-school operating systems were places where you installed and ran apps. You wanted to do something, you picked the right app, opened it, and clicked through steps. HarmonyOS 7 flips that. You state your intent — “help me sign up for a marathon” — and the system figures out which capabilities to call, then orchestrates them.

That's a big architectural shift, and it's powered by something called HMAF 2.0, the HarmonyOS intelligent agent framework. Liu breaks it into six layers: Xiaoyi (the system-level assistant), HMAF 2.0 (task decomposition and agent coordination), the AI base (openPangu 2.0 and a 30B on-device model), system-level guarantees (Ark engine, Star Shield security, and cross-device connectivity), developer tools (DevEco Code and DevEco CLI), and specific scenarios like spatial computing.

For developers, the most immediate change is that your app doesn't have to sit passively waiting to be opened. It can register itself as a schedulable agent. In the demo, Xiaoyi takes a single sentence, breaks it into subtasks, and dispatches them to sub-agents like “health,” “calendar,” and “search.” On the code side, that means exposing your app's capabilities through HMAF — declaring intents, parameters, and callbacks. It's a far cry from a voice assistant making a single API call.

Two tracks for developer tools

Huawei's tool strategy is what Liu calls a “dual-track” approach. DevEco Code is the full assistant — it has a brain, plans, writes code, compiles, and even fixes errors automatically. DevEco CLI, on the other hand, is all about openness. It exposes HarmonyOS's atomic capabilities (project management, build checks, debugging) as commands that any AI agent — Claude, Cursor, or your own in-house system — can call.

The two serve different audiences. DevEco Code is for new teams and greenfield projects. DevEco CLI is for teams that already have an AI agent pipeline and just want to plug HarmonyOS into it. They're complementary, not competitive.

Under the hood, DevEco Code is built on Huawei's self-developed Bifang engine plus the open-source OpenCode framework. Bifang handles the agent's thinking and tool use (think Claude Agent SDK). OpenCode provides the terminal UI, configuration, and the ecosystem of MCP, skills, and plugins. The split is deliberate: self-developed parts get deep integration with HarmonyOS tooling, while the open parts keep third-party tools compatible.

Real-world friction: fragmentation and missing corpus

For smaller teams, the biggest headache is device fragmentation. HarmonyOS runs on everything from flagship phones to cheap entry-level models, plus tablets, cars, and wearables. With limited test devices, bugs often surface only after an app goes live — install failures, startup crashes, layout glitches, or lag on certain hardware.

Huawei offers some help. EasyGo's parallel view lets you write one config file to get a proper landscape experience on foldables and tablets. And there's an automated UX checker that flags layout issues like oversized text or cut-off content, pointing straight to the source line. But Liu is candid about the gaps.

First, DevEco Code doesn't run on Linux — a pain for server-side and open-source workflows. Second, it's still tied closely to DevEco Studio, so pure command-line use is limited. Third — and most important — there just isn't enough ArkTS code in the training data for general-purpose AI models. Swift and Kotlin have years of corpus behind them; ArkTS is young. Liu says AI-generated ArkTS code needs manual fixes about 15–20% of the time.

The community is stepping in. Projects like harmonyos-ai-skill compress thousands of lines of HarmonyOS knowledge into a single Markdown file. Once you configure it, Claude, Cursor, and Copilot can tap into that knowledge, partially closing the gap.

How HarmonyOS compares with Apple and Google

At HDC, Huawei put developer tools and runtime AI in one keynote, presenting a unified “AI operating system” story. Apple and Google tell separate narratives — Xcode and Apple Intelligence are distinct; Google splits AI tools, models, and Android Studio across sessions. Liu sees this as more than presentation style: when a company draws one architecture diagram that includes both developer-side and runtime agents, it signals a single strategic push.

On architecture philosophy, Apple goes “open access” via MCP bridging, letting third-party agents plug into Xcode. Google shut down its open-source Gemini CLI in favor of a closed Antigravity, which ruffled feathers. Huawei is dual-track again: a full assistant for those who want it, and an open CLI for those who don't.

On pricing, Xcode is free but third-party models cost extra (Claude Pro starts around $20/month). Google's enterprise AI runs $45–$100 per user per month. Huawei gives you free access with GLM-5.1 built in, allowing 50 calls per minute, and lets you switch to DeepSeek or OpenAI models if you prefer. The message is clear: get developers in the door first, worry about revenue later.

Skill ecosystems differ too. Apple has a few official agent skills; Google goes cloud-first with no local skill count; Huawei offers 70+ curated skills covering multi-device development, issue diagnosis, and meta-service generation. Interestingly, both Apple and Huawei use the SKILL.md open format, which could become a de facto standard for “skill-as-code.”

Cross-device: the OS-level difference

HarmonyOS's distributed soft bus puts cross-device capability in the OS itself. You can move a video stream from your phone to a nearby smart screen with a single system call, and it works across brands. Android relies on a patchwork of protocols (Wear OS, Android Auto, Matter) with inconsistent experiences. Apple's Continuity is smooth but locked to Apple devices. In code, that means HarmonyOS lets you invoke a capability on another device almost synchronously, while Android often requires stitching together vendor-specific implementations.

When it comes to cross-platform UI, ArkUI is the only framework that truly runs on phones, tablets, PCs, cars, watches, and big screens with one codebase. Compose Multiplatform is still evolving, and SwiftUI is mature but Apple-only. Liu's verdict: HarmonyOS's differentiation here isn't marketing — it's baked into the OS.

Agent Loop: closing the loop between dev-time and runtime

HarmonyOS has 19% of the Chinese smartphone OS market (per Counterpoint), with 11 million registered developers and 400,000 apps — but only 23,000 have completed native adaptation. That gap is exactly what the AI tools aim to close.

Kuaishou is a case study. Using HarmonyOS AI tools, they hit 80% AI code generation, 84% acceptance of AI-generated test cases, and 73% acceptance of AI repair suggestions — a 1.7x overall efficiency boost. One engineer can now deliver a phone app, and two can handle phone, tablet, and car screens without extra HarmonyOS staff.

But Liu stresses the real lesson isn't the 80% number. Kuaishou's own Kwaipilot had already reached 30–40% code generation, yet delivery speed didn't improve. Writing code faster doesn't help if analysis, design, refactoring, and verification remain slow. So they worked with the HarmonyOS team to build a special skill called Ark Refiner-Sendable that automates the entire concurrency-safety refactoring loop. What used to take two people a week now takes half a day, and cold-start performance improved 16%.

The takeaway: investing in specialized skills for concrete engineering problems beats chasing raw code generation rates. It's an official case, so take it with a grain of salt, but the method is replicable.

What developers should do now

Liu offers three practical pieces of advice. First, decide early whether DevEco Code or DevEco CLI fits your team — new projects go with Code; existing systems plug into CLI. Second, check the 70+ existing skills; concurrency safety and other HarmonyOS-specific pain points already have solutions you can reuse. Third, use community knowledge packs like harmonyos-ai-skill to give your favorite AI tool a quick HarmonyOS boost.

HarmonyOS 7 has its rough edges — no Linux support, thin ArkTS corpus, still-maturing ecosystem. But if there's one thing to remember, it's this: the deep connection between development-time and runtime agents might be the thing that sets HarmonyOS apart from Apple and Google. When AI can not only write your app but also make it natively understandable by the OS, the way we build software changes at a fundamental level.