What HarmonyOS 7's Agent-First Design Means for Developers

The Big Shift: From Apps to Intent

When HarmonyOS 7 launched in 2026, the headlines focused on performance and flashy features. But developers, pay attention: the entire OS is now built around AI agents, not apps. That's a fundamental change in how you'll think about building software.

Think about your daily phone habits. You want to book a flight, order food, or track a run—so you hunt for the right app, open it, and click through its menus. HarmonyOS 7 flips that. You just say what you need, and the system figures out which capabilities to use and combines them on the fly.

That's not a minor tweak. It's a whole new paradigm for an operating system.

HMAF 2.0: The Six-Layer Architecture

The heart of this shift is HMAF 2.0, Huawei's upgraded agent framework. It's structured in six layers:

• XiaoYi – the system-level assistant, your entry point
• HMAF 2.0 – breaks down tasks and coordinates multiple agents
• AI base – includes the open-source openPangu 2.0 and a 30B on-device model
• System services – Ark engine, StarShield security, and Galaxy Connect
• Developer tools – DevEco Code and DevEco CLI
• Scenarios – like spatial computing

For developers, the key takeaway is that your app can now expose itself as a schedulable agent. In the keynote, Huawei showed a demo where a user said “help me register for a marathon,” and the system split that into subtasks, coordinating health, calendar, and search agents simultaneously.

Under the hood, it's about registering an agent with HMAF. You declare what your app can do, define the input schema, and provide a callback. When the system matches a user's intent to your capability, it hands you a structured task—not a vague natural language query.

openPangu 2.0 and Performance Gains

Huawei also open-sourced openPangu 2.0, with a Pro version at 505 billion parameters and a Flash version at 92 billion. Both support 512K context. But the more intriguing claim is Ascend-native optimization—single-card throughput supposedly twice that of mainstream open models.

Performance-wise, HarmonyOS 7 introduces a performance model in the scheduler. App startup is 24% faster for system apps, 34% for ecosystem apps, and game frame rate stability improves 40%. And they've managed to keep annual load growth under 10%.

Security gets an AI boost too. The StarShield architecture uses on-device AI to detect scams, identifying seven major fraud patterns in seconds. Huawei claims it has already prevented 3.47 million potential scams, with apps like Alipay and Douyin onboard.

DevEco Code and DevEco CLI: A Two-Track Strategy

Huawei's developer tools now follow a dual-track approach.

DevEco Code is the co-pilot—it has its own brain. You give it a requirement, and it plans, writes, compiles, and debugs, even fixing errors automatically. It's built on Huawei's self-developed Bifang engine (which handles agent reasoning and planning) plus the open-source OpenCode framework (which handles terminal interaction and ecosystem hooks).

Two agents work inside DevEco Code: a Plan Agent that breaks down requirements into a plan, and a Build Agent that executes—writing code, compiling, and fixing issues. It even modifies resources and pages in your project. For instance, if you need your app to adapt to phones and tablets, the Plan Agent inserts breakpoint-based layouts automatically.

DevEco CLI is the opposite—it doesn't think. It just exposes Huawei's atomic capabilities (project management, build checks, debugging) as commands. That means you can plug it into whatever agent you already use—Claude, Cursor, or your own CI system.

The two tracks serve different crowds. DevEco Code is for new teams and projects. DevEco CLI is for big teams with existing agent pipelines who don't want to change their workflow.

Real-World Pain Points: Fragmentation and Beyond

For smaller teams, the biggest headache is device fragmentation. HarmonyOS runs on everything from budget phones to flagships, plus tablets, cars, and wearables. Each has different screen sizes, chips, memory, and API levels. Most small teams only have a few test devices, so issues like install failures, crashes, or layout glitches often surface only after users download the app.

Huawei offers some help. EasyGo Parallel View lets you write a config file to get a landscape experience on foldables and tablets. There's also an automated UX detection tool that flags layout problems and points to the exact source line. But there are gaps.

First, DevEco Code doesn't support Linux—only Windows and macOS. That's a problem for open-source and server-side devs. Second, it's still tied to DevEco Studio; the pure command-line experience is limited. Third, and most impactful, is the lack of ArkTS training data in general-purpose AI models. When you ask GPT or Claude to write Swift or Kotlin, it works out of the box. With ArkTS, you might need to hand-fix 15–20% of the generated code.

The community is stepping up. Projects like harmonyos-ai-skill compress thousands of lines of HarmonyOS knowledge into a single Markdown file. Once configured, Claude, Cursor, and Copilot can all tap into it.

How HarmonyOS Differs from iOS and Android

At HDC, Apple and Google also talked about AI dev tools, but their framing was different. Apple kept Xcode and Apple Intelligence as separate stories. Google split its keynote into four parts. Huawei, by contrast, put developer tools and runtime AI into one big architecture diagram.

That's not just presentation. It signals a strategic choice: Huawei is treating development-time and runtime agents as one system.

Looking at the specifics:

• Architecture philosophy: Apple uses an open-access approach with MCP bridging; Google moved from open-source Gemini CLI to closed-source Antigravity; Huawei runs a dual-track.
• Model strategy: Apple lets third-party models compete (you pay for them); Google bundles Gemini with enterprise pricing starting at $45/user/month; Huawei offers free use with built-in GLM-5.1 (50 calls/min) and lets you switch to DeepSeek or OpenAI.
• Skill ecosystem: Apple has a few in-house skills; Google relies on cloud; Huawei offers 70+ curated skills covering multi-device development, problem diagnosis, and meta-service generation.

One detail worth noting: both Apple and Huawei use the SKILL.md open format, which suggests Skill-as-Code is becoming a de facto standard.

Under the hood, cross-device connectivity also differs. HarmonyOS's distributed bus is system-level, enabling plug-and-play across brands. Android cobbles together Wear OS, Auto, and Matter. Apple's Continuity is smooth but only within its own ecosystem.

Can Development-Time and Runtime Agents Truly Connect?

HarmonyOS's market share in China hit 19% in May, surpassing iOS for seven straight quarters. There are over 11 million registered developers and 400,000 apps and services. But only 23,000 are natively adapted. The gap isn't developer interest—it's the effort required to adapt.

AI tools aim to close that gap. Kuaishou, a real production case, reported 80% AI code generation in development, 84% direct adoption of AI-generated test cases, and 73% adoption of AI fix suggestions. Two engineers, using these tools, could deliver phone, tablet, and car versions simultaneously—previously, one engineer handled just a phone.

The key insight from Kuaishou's case isn't the 80% number. It's that they built a custom skill for HarmonyOS concurrency safety (Ark Refiner-Sendable) that automated the whole analysis-to-verification loop. A task that took two people a week now takes half a day, with a 16% improvement in cold start.

That's the real promise: not just generating code faster, but solving specific engineering problems with reusable skills.

What Developers Should Do Next

If you're diving into HarmonyOS, here are three practical steps:

1. Choose your track early. New project? Go with DevEco Code. Existing system? Use DevEco CLI to integrate without rewriting everything.
2. Leverage the 70+ existing skills. Problems like concurrency safety already have ready-made solutions. Don't reinvent them.
3. Use community knowledge packs. A one-time setup can give your favorite AI tools HarmonyOS knowledge at low cost.

HarmonyOS 7 isn't perfect—Linux support is missing, training data is thin, and the ecosystem is still maturing. But the direction is clear: the deep integration of development-time and runtime agents is what sets it apart from Apple and Google. For developers, that could change not just how you write code, but what you build with it.