Google向第三方 AI agents 和发布新的 Gemini API 打开 Android 开发
关键点
- Google正通过"Bring Your Own Agent"功能，将其Android Studio开发环境向Claude和Codex等40多个外部编码agents开放，同时扩展终端工具以减少开发时间和象征性使用。
- 使用Gemini Nano 4动力Agentic API和模型上下文协议等接口，开发者可以在移动设备上局部运行agent功能，并将应用服务直接插入Google搜索和Gemini应用。
- Google和硬件制造器XREAL还发布是XREAL AURA，是一款运行于Android XR的1 099双智能眼镜，显示虚拟工作空间，并使用Gemini分析现实世界环境。
Google正围绕AI agents在Android Dev峰会上以开发环境为中心。 更新内容包括Anthropic和OpenAI型号的集成，本地智能手机agents的新API，以及一对智能眼镜。
Google 推出在今年的Android Dev峰会上为其开发工具和系统API更新。 开放Android Studio开发环境是主要重点。
与Android Studio Rabbit 2,Google为发布，是一个"Bring your Own Agent"(BYOA)功能。 开发者不再局限于Google自己的AI工具。 相反，他们可以使用开放的Agent客户协议(ACP)在IDE直接运行超过40个不同的编码agents.
开发者可以通过内置市场安装并同时运行Anthropic的Claude,OpenAI的代码，Google抗重力。 这些agents可以直接访问项目结构，符号分辨率，模拟器，以及Logcat系统等IDE工具。
Google还用Android座椅领导板的发布版本2.0测量复杂任务的AI模型性能。 它根据功能、时间和成本评价多日发展项目的语言模式。
AI生成的代码变化经常会变得混乱，因此Google正在增加一个名为"改变评论"的代码验证工具。 它分析待修改，生成错误报告，并一步步解释修改。 Android Studio Canary 预览版也是新增一个迁移助理，自动将较旧的XML布局翻译为 Jetpack Compose. 同时，Google正在测试Lightbuild，一个基于YAML schema的宣告性构建系统，人类和AI agents都可以阅读和编辑。
新的终端工具和专门技能用于编码 agents
Google扩展了Android CLI命令线工具，用于在终端或CI管道中运行其agents的开发者。 现在作为Claude Code,OpenAI的Codex和Google抗重力的官方agent插件可供使用。
新的命令让 AI agents 在终端中渲染 Compose 预览，在不启动模拟器的情况下直观检查 UI 更改。 内部的Google测试显示，agents完成UI任务的速度快了77%，并用到了这个特性的3倍以下的代币。 CLI还支持来自Google实验室的流体物理测试设备，因此agents可以调查设备特定的bug.
调试和性能分析也有新功能。 Google在CLI和Android Studio配置器中都建了开源漏气内存漏漏检测库。 系统评价电脑上的堆积堆积，而不是移动设备，这使得应用在测试时不会减速。 在使用Anthropic的Claude Opus 5.5模型的Google A/B测试中，CLI工具帮助查找和修复内存泄漏的速度几乎是快的两倍，同时将符号使用量削减了约25%.
其他命令行工具让agents捕捉系统通过Android性能分析仪跟踪，从Firebase崩溃分析仪中提取崩溃报告，或通过R8编译器检查代码优化。 语言模型往往缺乏最新API的培训，因此Google现在提供超过30个Android技能。 这些向agents提供文档，用于在应用程序提交前检查Google Play政策，处理许可安全问题，或调整以Googlebook为品牌的可折叠器和笔记本电脑的布局。
Gemini Nano 4 和 MCP 将 Android apps 变成交互式的 agent 工具
除开发者工具外，Google正在传输新的API，将agent特性构建为app或连接到系统agents的app. 对于本地的端侧处理，Google正在在其ML Kit中添加由Gemini Nano 4提供动力的α Agentic API. 此界面会自动执行函数，保存聊天历史，管理管弦乐圈，并提供安全过滤器。
Google正在使用模式背景协议(MCP)和发布，在β中的AppFunctions API作为统一的工具调用层，因此app可以与本地或全系统的agents共享功能。 Jetpack Compose 的新α Agent-to-UI(A2UI)接口以此为基础。 agents不使用纯文本块进行回复，而是可以显示由特定app组件构建的交互式用户界面。
为了将这些应用服务带入Google搜索和Gemini应用，公司是发布，是连接Apps平台的预览。 开发者可以通过Android MCP或云托管的MCP服务器链接自己的应用，并通过一个单一控制台管理。 Google 称来自Adobe,Peloton,Experian等公司的超过50个合作伙伴集成已经生活在Gemini的app中，它们将会在未来几周推出网络搜索。
XREAL AURA智能眼镜将Android XR和Gemini带入现实世界
Google还针对几个硬件形式因素进行了宣布更新。 更新至Jetpack Compose 1.13和App State模式使得应用更便于大屏幕的应用，扩展现实(XR)开发环境正在离开测试阶段。 Jetpack XR SDK,Android Studio中的XR仿真器，以及团结，哥多特和不真实的引擎枢纽，现在作为稳定的发布提供。
在软件更新的同时，Google与硬件制造者XREAL合作，为XREAL AURA揭幕。 XR队长Shahram Izadi 称他们是第一个运行着Android XR的wired智能眼镜。 眼镜重量在95克以下，并具有透明，可淡化的镜头，有70度的视野。 计算单元和电池坐落于翻倍作为轨迹盘的外部板块中，运行在Qualcomm新的Snapdragon Reality Elite芯片上。
用户可以一次将最多5个Android的应用程序放置在自己的视野领域，或者插上USB-C电缆来反射他们的笔记本电脑屏幕。 Gemini也是在这里建造的。 AI可以通过眼镜和虚拟屏幕看到现实世界的环境，使其可以组织windows，提供电子游戏提示，或者在房间里视觉突出物理物体。 XREAL AURA现在可以在美国、加拿大和韩国订购1 099美元，11月中旬开始运输。
AI News Without the Hype – 由人类治愈
订阅THE DECODER免费阅读，每周AI通讯，我们独家的"AI雷达"前沿报告每年6次，全面存档访问，访问我们的评论部分。
现在订阅
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Google opens Android development to third-party AI agents and unveils new Gemini APIs
Key Points
- Google is opening its Android Studio development environment to more than 40 external coding agents like Claude and Codex via the "Bring Your Own Agent" feature, while expanding terminal tools to cut down on development time and token usage.
- Using interfaces like the Gemini Nano 4-powered Agentic API and the Model Context Protocol, developers can run agent functions locally on mobile devices and plug app services directly into Google Search and the Gemini app.
- Google and hardware maker XREAL also unveiled the XREAL AURA, a $1,099 pair of smart glasses running Android XR that displays virtual workspaces and uses Gemini to analyze the real-world environment.
Google is centering its development environment around AI agents at the Android Dev Summit. The updates include integrations for Anthropic and OpenAI models, new APIs for local smartphone agents, and a pair of smart glasses.
Google introduced updates for its developer tools and system APIs at this year's Android Dev Summit. Opening up the Android Studio development environment was the main focus.
With Android Studio Rabbit 2, Google is launching a "Bring Your Own Agent" (BYOA) feature. Developers are no longer limited to Google's own AI tools. Instead, they can use the open Agent Client Protocol (ACP) to run more than 40 different coding agents directly in the IDE.
Developers can install and simultaneously run Anthropic's Claude Agent, OpenAI's Codex, and Google Antigravity through the built-in marketplace. These agents get direct access to IDE tools like project structure, symbol resolution, the emulator, and the Logcat system.
Google also released version 2.0 of the Android Bench leaderboard to measure AI model performance on complex tasks. It evaluates language models on multi-day development projects based on functionality, time spent, and cost.
AI-generated code changes often get messy, so Google is adding a code verification tool called Change Review. It analyzes pending changes, generates bug reports, and explains modifications step by step. The Android Studio Canary preview also adds a migration assistant that automatically translates older XML layouts into Jetpack Compose. Meanwhile, Google is testing Lightbuild, a declarative build system based on a YAML schema that both humans and AI agents can read and edit.
New terminal tools and specialized skills cut token usage for coding agents
Google expanded the Android CLI command-line tool for developers who run their agents in the terminal or in CI pipelines. It is now available as an official agent plugin for Claude Code, OpenAI's Codex, and Google Antigravity.
New commands let AI agents render Compose previews in the terminal to visually check UI changes without booting up an emulator. Internal Google tests show agents finished UI tasks 77 percent faster and used up to three times fewer tokens with this feature. The CLI also supports streaming physical test devices from Google labs so agents can investigate device-specific bugs.
Debugging and performance analysis got new features too. Google built the open-source LeakCanary memory leak detection library into both the CLI and the Android Studio Profiler. The system evaluates heap dumps on the computer instead of the mobile device, which keeps the app from slowing down during testing. In a Google A/B test using Anthropic's Claude Opus 5.5 model, the CLI tool helped find and fix memory leaks nearly twice as fast while cutting token usage by about 25 percent.
Other command-line tools let agents capture system traces via the Android Performance Analyzer, pull crash reports from Firebase Crashlytics, or check code optimization through the R8 compiler. Language models often lack training on the newest APIs, so Google now provides over 30 Android Skills. These feed documentation to agents for tasks like checking Google Play policies before app submission, handling permission security issues, or adapting layouts for foldables and laptops branded as Googlebooks.
Gemini Nano 4 and MCP turn Android apps into interactive agent tools
Beyond developer tools, Google is shipping new APIs to build agent features into apps or connect apps to system agents. For local on-device processing, Google is adding an alpha Agentic API powered by Gemini Nano 4 to its ML Kit. This interface automatically executes functions, saves chat history, manages the orchestration loop, and provides safety filters.
Google is using the Model Context Protocol (MCP) and launching the AppFunctions API in beta as a unified tool-calling layer so apps can share functions with local or system-wide agents. The new alpha Agent-to-UI (A2UI) interface for Jetpack Compose builds on this. Instead of replying with plain text blocks, agents can display interactive user interfaces built from the specific app's components.
To bring these app services into Google Search and the Gemini app, the company is launching a preview of the Connected Apps Platform. Developers can link their apps via Android MCP or cloud-hosted MCP servers and manage them through a single console. Google says over 50 partner integrations from companies like Adobe, Peloton, and Experian are already live in the Gemini app, and they will roll out to web search in the coming weeks.
XREAL AURA smart glasses bring Android XR and Gemini into the real world
Google also announced updates for several hardware form factors. Updates to Jetpack Compose 1.13 and the AppState model make it easier to adapt apps for large screens, and the Extended Reality (XR) development environment is leaving the testing phase. The Jetpack XR SDK, the XR emulators in Android Studio, and the Engine Hub for Unity, Godot, and Unreal are now available as stable releases.
Alongside the software updates, Google partnered with hardware maker XREAL to unveil the XREAL AURA. XR chief Shahram Izadi says they are the first wired smart glasses running Android XR. The glasses weigh under 95 grams and feature transparent, dimmable lenses with a 70-degree field of view. The computing unit and battery sit in an external puck that doubles as a trackpad and runs on Qualcomm's new Snapdragon Reality Elite chip.
Users can place up to five Android apps in their field of view at once or plug in a USB-C cable to mirror their laptop screen. Gemini is built in here too. The AI can see both the real-world environment through the glasses and the virtual screens, allowing it to organize windows, offer video game tips, or visually highlight physical objects in the room. The XREAL AURA is available to order now in the US, Canada, and South Korea for $1,099, with shipping set to begin in mid-November.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
Auto-fetched and translated by AI Atlas. All rights belong to the original author.