HarmonyOS AI Coding: A New R&D Paradigm and Engineering Practice

Summary

This article presents Huawei’s engineering approach to AI-assisted HarmonyOS development and explains why the ecosystem faces challenges beyond those of Android and iOS. ArkTS has a shorter history and less publicly available code than Java, Kotlin, Swift, and Objective-C, while HarmonyOS must also support a wider range of device forms, including straight-screen phones, foldables, and tablets. The proposed toolchain has three layers: DevEco Code is an out-of-the-box Coding Agent covering project creation, coding, testing, and maintenance; DevEco CLI breaks IDE functions into programmable command-line tools; and the Gewu marketplace aggregates Skills, MCP tools, and knowledge bases. DevEco Code uses a Harness with syntax and dependency checks, compilation diagnostics, multimodal UI-intent verification, and automated repair. Its UI agent launches an emulator, performs actions such as tapping, swiping, and text input, captures screenshots, and returns visual comparison evidence. The article says the repair agent covers more than ten classes of errors and achieved an over-80% success rate on runtime-crash repair based on more than 100 accumulated cases. It also describes Plan, Goal, and Build modes, which provide different levels of autonomy and human confirmation. DevEco CLI includes a 20-million-character official HarmonyOS knowledge base, more than 70 Skills, and support for initialization and integration with tools such as OpenCode. A four-level Skill-routing design directs agents from ArkTS syntax and system APIs to project context and business rules, reducing context overload. HarmonyOS Verify adds compile checks, UI automation, visual comparison, diagnostic tools, and multi-device validation. The article reports that Harness improved task completion over OpenCode across combinations using GLM5.1, DeepSeek-V4-Pro, and MiniMax-M2.7, while identifying high-quality HarmonyOS code data, large-project understanding, and long-horizon task control as remaining ecosystem challenges.